Imagine you’re a freelance designer waiting for a client to approve a website. You keep opening the same email thread because one reply could change tomorrow’s work. Most visits tell you nothing. One eventually matters: the client wants to launch on Thursday instead of next Tuesday.

An AI assistant could help by watching for that change. It could also send you a cheerful summary of every message, every morning, and leave you to work out which one matters. Now you have the original thread and a briefing to check. Congratulations on the extra inbox.

This is the question behind the new wave of always-on assistants: what should they bring to your attention, and what should they leave alone?

OpenAI introduced dots last week, personal AI agents with their own cloud computers that can work between conversations. The interesting part is the possibility of useful work starting before you remember to ask. That also means choosing when an assistant gets to interrupt you.

📌 TL;DR

• An assistant needs to know which changes would affect your next decision. “Keep me updated” leaves that judgment almost completely open.

• A useful alert brings the source, the consequence and something prepared for your review. A summary alone may leave the hard part untouched.

• Silence only helps when you can verify that the checking worked. Review missed events as well as unnecessary alerts.

• Start with one low-stakes source and one kind of change. Use real examples to decide whether the assistant has earned a recurring job.

A new message isn’t always a new decision

Return to the website project. The client might reply “looks good,” ask whether the blue can be slightly darker, or move the launch forward by several days. All three are new messages. They don’t deserve the same response.

A vague request to “monitor the project and flag anything important” hands the assistant a word it can’t apply well without more context. Important to whom? A designer waiting for approval cares about different changes from a bookkeeper waiting for an invoice number. The same message can be routine for one person and urgent for another.

A better instruction begins with the decision you’re trying to protect. In this example: “Tell me when client feedback changes the agreed launch date, adds work that could affect that date, or prevents me from continuing.” The assistant now has a reason to bring something forward. It can connect the message to the work it changes.

This doesn’t require predicting every possible message. Start with a few examples. A thank-you can wait. A new page requested before launch needs attention. “Could we have it earlier?” needs clarification because there is no new date to plan around yet. The last example matters: uncertainty can deserve attention without being treated as a confirmed change.

You can also tell the assistant which interruptions you’ve already decided against. If a color revision belongs in the next scheduled review, a new comment about blue should not automatically produce a separate alert. Otherwise, an assistant that is technically following the conversation can make it harder to finish the design.

Dots make this question timely because they can do ongoing work rather than wait for each individual prompt. But “ongoing” doesn’t establish how often a particular source is checked. A scheduled check, an event from a supported app and occasional background research are different ways of finding something. None should be assumed from a friendly promise to keep an eye on it.

Before relying on a routine, ask how it will run and inspect the actual setup. If a missed message would create a serious problem, keep the existing dependable alert while testing the AI. The purpose of the first experiment is to learn whether its judgment is useful, not to discover the hard way that a connection wasn’t working.

The practical improvement is quite small: replace “tell me what’s new” with “tell me what would change this decision.” That gives you something specific to judge when the first alert arrives.

The interruption should arrive with useful work

Suppose the designer receives this alert: “The client has requested an earlier launch.” It found the relevant message. You now have to reopen the thread, locate the date, compare it with the remaining work and decide what to ask next.

A more useful alert would say that the client proposed Thursday, link to the message and point out that the agreed plan still includes two unfinished pages. It might prepare a reply asking whether those pages can follow after launch. You can inspect the evidence and decide whether that trade works.

Notice where the value comes from. The assistant has brought together facts that would otherwise require another round of checking. Its proposed reply is helpful because it addresses the actual conflict. A confident recommendation without those facts would just give you another claim to investigate.

OpenAI’s launch article describes an early user whose dot spotted a forgotten invoice connected to a publication. According to OpenAI, the dot prepared the invoice and sent it after the user approved. That’s a vendor-reported example, not an independent measure of how reliably dots find unpaid work. But it illustrates a useful standard: the discovery led to something the person could review and finish.

There is an important product distinction here. OpenAI’s privacy and safety FAQ says proactive research is read-only. It cannot send messages, change connected apps or use the browser or computer on its own. Follow-up actions have separate permission checks. Finding something worth doing and having authority to do it are separate questions.

For the website example, the first routine can stay modest. Read the chosen thread, identify the relevant change and draft a question. Sending that question or agreeing to Thursday would be a further action. There is no reason to bundle those permissions together just to find out whether the assistant is good at noticing the change.

The alert should also distinguish what the source says from what the assistant concludes. “The client asked for Thursday” is a fact you can check in the message. “The remaining work may not fit” is an assessment that depends on the work estimate. “Offer a smaller launch” is a proposal. Keeping those parts clear makes disagreement easier.

If the assistant has missed a completed page, you can correct the project facts. If its time estimate is wrong, you can correct the estimate. If you don’t want to offer a smaller launch, you can reject the proposal without throwing away the useful discovery.

That’s the standard worth applying to a proactive assistant: when it asks for your attention, does it make the next decision easier? The answer should be visible in the work it brings, rather than in how convincingly it explains its enthusiasm for helping.

Silence needs a way to check it

There is one uncomfortable part of this setup. An assistant that sends too many alerts is visibly annoying. An assistant that misses the important message may look wonderfully unobtrusive.

You therefore need to check two kinds of mistake. Did it interrupt you about something that could wait? And did it fail to surface something that should have changed your plan? Reading only the alerts can answer the first question. It cannot answer the second.

During the first few runs, compare the assistant’s output with the original source. In the designer example, look through the client thread and check whether each meaningful change was picked up. This is extra work during the trial. If maintaining confidence continues to require rereading everything indefinitely, that cost belongs in the decision about whether to keep the routine.

Ask for a brief record of what was checked and when, separate from the notifications you receive. “No relevant change found in the client thread checked this morning” gives you something to inspect. “Couldn’t access the thread” is a different result and must not disappear into silence. An access failure should be reported because the assistant can no longer perform the job you are relying on.

That record isn’t proof that it read everything correctly. It makes the checking inspectable. You can compare its account with the source, find a missed reply and then work out whether the problem was access, timing or judgment.

Those failures need different fixes. Rewriting “be more careful” won’t restore a broken connection. Checking more often won’t teach it why a vague request matters. And a stricter filter could reduce interruptions while making the important misses worse.

Use an actual mistake to adjust the smallest relevant part. If “could we have it earlier?” went unnoticed, add it as an example of a message that needs a question. If every minor visual revision produces an alert, explain which revisions belong in the scheduled review. Then check the next results before expanding the job.

A routine is worth keeping when the useful discoveries justify the attention and checking it requires. There is no universal right number of alerts. One unnecessary interruption may be acceptable for a time-sensitive client decision and excessive for a casual reading list. Set the standard around the consequence you’re trying to avoid.

Also this week

GPT-6.1 Sol arrived in Work and Codex. OpenAI positions the new model as a lower-cost option for a wide range of work, approaching Astra in its evaluations. That is a reason to compare it on a familiar task, not to assume every difficult job will come out the same. The launch rollout covers Work and Codex rather than regular ChatGPT chat.

Claude Sonnet 5.5 targets everyday work. Anthropic’s September 28 release emphasizes well-scoped tasks, including documents, slides and spreadsheets. The company reports faster responses and lower costs per task than Sonnet 5, while saying Opus remains stronger for complex, open-ended judgment. Treat that distinction as a starting point for a comparison on your own work, rather than a universal model ranking.

Meta is bringing Muse to small businesses. Its September 29 announcement describes an assistant that works across business apps, with approval required for sending, publishing and spending. Access is limited, including age and regional restrictions, so check eligibility before assuming it’s available. The practical question is familiar: which recurring job would it complete well enough to justify another tool?

Try one small watch before adding ten

Choose a repeated check you already make, where a mistake during a trial would be manageable. A client thread you will continue reading, a shared project document or a small reading list is enough. Avoid making the first test responsible for something you can’t afford to miss.

Before scheduling anything, give the assistant a handful of old examples from that source and ask what it would have brought to your attention. Include one routine update, one clear change and one ambiguous message. Explain your decisions where they differ. This lets you inspect its interpretation before it starts producing recurring alerts.

Here’s a sample request to adapt. It is a suggested starting point, not a tested setup or a promise that every assistant supports the same connections:

Review this client thread for changes to the agreed launch date, added work that could affect it, or questions that prevent the next task from moving forward. For each relevant change, link to the message, explain what it could change in the plan and draft the next question if one is needed. Distinguish confirmed changes from requests that need clarification. Don’t send anything. Treat routine acknowledgments and minor visual revisions as items for the next scheduled review. Tell me if you can’t access the source.

If those example judgments are useful, set one recurring check at a time that suits the work. Verify the selected source, schedule and notification settings. In dots, the getting-started guide explains how to request recurring work and inspect scheduled tasks and recent activity. Check the saved routine rather than treating the assistant’s conversational acknowledgment as the entire setup.

Decide when you’ll review the trial. After several real checks, look at the useful alerts, the interruptions you didn’t need and anything it missed. Keep it, change the instructions or stop it based on that evidence. Don’t add five more sources because the first briefing looked impressive.

🧰 Tools to try

Dots, if your account has access. This is a relevant place to try the recurring-work idea, but availability is limited. OpenAI’s current guide lists Pro access outside the EEA, Switzerland and the UK, Business Premium in supported ChatGPT regions, and an admin-enabled beta for certain organization plans. Rollout can vary. Check access before planning a workflow around it; this issue isn’t a reason to buy a new subscription.

The alerts in the app you already use. If the rule is simply “notify me when this person replies,” an ordinary email or app notification may cover the job. AI becomes more interesting when the question requires interpretation, such as whether a request changes the agreed workload. Match the tool to that missing judgment before adding another system to maintain.

🥣 Brain food

OpenAI’s explanation of dots’ safety design is worth reading with a specific question in mind: which parts of a proposed routine are limited by the tools, and which depend on the model making a good decision? The document describes restrictions on proactive research and controls for subsequent actions. Those controls don’t establish whether an alert will matter to you. That is what your examples and trial are for.

Return to the designer waiting for a reply. The useful outcome is a clear notice that Thursday changes the plan, with the message and a sensible question ready to review. A longer daily digest doesn’t improve that outcome by being longer.

If the assistant can reliably bring forward the changes that matter, you have a reason to stop reopening the same thread quite so often. If you must check every message and then inspect its summary too, it hasn’t earned that trust yet. Let the routine prove useful before giving it more of your attention.

P.S. An AI wrote these sentences from this week’s primary sources. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.

An AI can work through a project with you all afternoon and still miss an important detail when you return tomorrow. It can also answer a follow-up faster because part of the earlier work was reused. Those two experiences seem contradictory if you describe both as “memory.” They make more sense once you separate the information the assistant receives from the computation used to process it.

OpenAI’s September 22 update to prompt caching gives us a useful way into that distinction. The announcement concerns the cost of repeatedly processing familiar material. It does not promise that every detail you tell an assistant will become a reliable fact it can recall forever.

You do not need to become a developer to understand the difference. It explains something practical: when an assistant keeps using an old detail, or loses one you expected it to keep, asking it to “remember better” may leave the actual problem untouched.

📌 TL;DR

• Repeated background can become cheaper → Prompt caching lets an AI service reuse computation from earlier input. The discount applies to eligible reused input, not automatically to your whole bill or subscription.

• Reuse and recall are different → A temporary cache, the current conversation, and a product’s memory features do different jobs. One can work while another disappoints you.

• Correct the source of the detail → A stale project file needs a different correction from an outdated personal preference. Find where the information came from before deciding what to change.

Repeating the background should cost less than starting over

Suppose you are using an assistant to plan a small garden. You provide a description of the beds, where the shade falls, and the plants already there. First you ask about the layout. Then you ask about a path. Then you change your mind about the seating area.

The later questions still depend on the original garden. A useful answer about the path needs to account for the beds, and a useful answer about the seating needs to account for the shade. The short new message is only part of the material needed to answer.

Behind many AI applications, software sends a request to the model through an API, which is the connection that lets one program use another service. That request can include earlier messages, background documents, and instructions as well as your newest question. The model’s input is therefore larger than the sentence you just typed.

OpenAI’s September 22 caching announcement describes reusing computation for that shared material. If the beginning of a later request matches eligible earlier input, the service can reuse work instead of processing all of it afresh. OpenAI says its improved system raises cache reuse by default and offers discounts of up to 90 percent on cached input.

That is a meaningful improvement for applications that keep carrying substantial background through a series of steps. It can reduce the repeated cost of supplying instructions while an agent works. It does not mean the agent has learned those instructions permanently, or that it can recover them from anywhere later.

There is a small piece of arithmetic worth understanding before the percentage runs away with the story. OpenAI’s current developer guide lists a cache-write cost of 1.25 times the ordinary input rate for GPT-5.6 and later, followed by a read cost of 0.1 times that rate.

Imagine processing one unchanged block would normally cost one unit. Processing it ten times without reuse would cost ten units. Writing it to the cache once and successfully reusing it nine times would cost 1.25 plus nine lots of 0.1, or 2.15 units.

This is an illustrative calculation for that input block alone. It assumes every later request successfully reuses the whole block. New material and generated answers still have their own costs. A 90 percent discount on a cached read therefore does not make the entire ten-step job 90 percent cheaper, and it says nothing about a reduction in your monthly subscription price.

Anthropic’s caching documentation describes the same broad idea of reusing a matching prompt prefix, with its own model prices and cache lifetimes. The details differ between services. What carries across is the reason for doing it: repeated background does not always need to incur the same processing cost each time.

If you use an ordinary subscription, the application handles those mechanics for you. There is no reason to reorganize every sentence you type around an API pricing table. The useful news is that providers are improving the economics of continuing work, alongside the more visible competition to make models smarter.

A cache does not decide what tomorrow’s chat remembers

Return to the garden. During today’s conversation, the assistant has been told that the back fence faces north. Tomorrow you open a different chat and ask where to put the bench. Whether that new answer has access to the fence detail is a question about how the application supplies information.

A cache cannot, by itself, decide that this particular sentence belongs in tomorrow’s planning conversation. It speeds the processing of matching material presented to the model. The application still needs a way to carry forward or retrieve relevant information.

OpenAI’s conversation-state guide explains ways developers can preserve conversational context, including supplying previous messages and using stored conversation state. That is a separate responsibility from making repeated input cheaper. A program needs both a sensible selection of material and a way to process it economically if it is going to support sustained work well.

You can see why these functions should remain distinct. The garden plan might contain several abandoned ideas. You considered putting a bench beside the fence, rejected it, and chose a shaded corner instead. Keeping every sentence available is useful history, but availability alone does not tell the assistant which decision is current.

Now imagine that somebody supplied only the first half of the discussion to the next session. The assistant could process that material efficiently and still recommend the rejected location. The missing correction would be an information problem. Faster processing would simply get you the outdated suggestion sooner.

This is a hypothetical example, not a report of a product test. It illustrates why “the AI forgot” can describe several different failures. Perhaps the information was never carried forward. Perhaps an older source contradicted it. Perhaps the right material was present and the model interpreted it incorrectly. Those possibilities deserve different investigations.

Product memory adds another function. OpenAI’s Memory guide says ChatGPT can use relevant preferences and details from available sources, depending on the account and its settings. It also says Memory does not retain every detail from every conversation. A remembered preference can help personalize an answer without serving as a complete record of a complicated project.

For example, “I prefer low-maintenance plants” is a useful general preference. “We rejected the bench by the fence because it blocks access to the shed” is a specific project decision. Both could matter to a garden answer, but they are different kinds of information. You would not want a passing comment about a friend’s garden quietly replacing the dimensions of your own.

Nor should you treat a cache lifetime as a privacy countdown. The current OpenAI caching guide describes a 30-minute eligibility window after a cache write or reuse for GPT-5.6 and later, and says entries may be retained longer. That is not a promise that your chat, files, or saved memories disappear after half an hour.

A familiar answer, a quick answer, and a correctly remembered answer can look similar from the outside. The difference becomes visible when you ask which information supported the result. That is a more useful question than judging the assistant’s memory by how confidently it speaks about your garden.

Correct the source that supplied the wrong detail

Suppose the assistant keeps describing the garden as having three beds, although you now have two. You could repeat “two beds” in every message. That may help the next answer, but it leaves you doing the correction again whenever the old description returns.

Start by locating the old detail. Is it in the file you uploaded? In the instructions attached to the project? In a previous conversation the product can reference? Or is the assistant making an unsupported assumption? Asking for the supporting source gives you something specific to inspect, though the assistant’s explanation still needs checking against the material itself.

If the uploaded plan shows three beds, update the plan you intend to use. If the file is correct and the assistant misread it, point to the relevant passage. If the answer provides no support, ask it to leave the number unresolved until it has evidence. These corrections are more precise than treating every wrong detail as a broken memory setting.

A temporary instruction needs care too. Imagine asking for a design with no flowers because you are exploring what a mostly green garden might look like. That does not necessarily mean you have developed a permanent dislike of flowers. Saying “for this version” makes the scope clearer and avoids turning an experiment into a standing preference.

The reverse happens when a real change gets phrased as a casual aside. “We removed the third bed last weekend” changes the current plan. A useful follow-up would identify the old plan as outdated and supply the replacement, so the assistant has a clear reason to use the newer dimensions. The point is to make the relationship between sources understandable.

ChatGPT’s Projects documentation describes keeping related chats, files, and instructions together. It also distinguishes default memory from project-only memory. In the latter mode, chats can draw on other conversations inside that project while excluding outside conversations and previously saved personal memories. The current guide also says ChatGPT Work is unavailable in a project using project-only memory, so this is a choice with a functional tradeoff.

That can explain an apparent inconsistency without blaming the model. A preference you established elsewhere may be outside the context that project is allowed to use. If you want that preference to guide the project, provide the relevant instruction there. Check the settings and features your account actually offers rather than assuming every project draws from everything you have ever said.

There is no prize for collecting the largest possible pile of context. A garden assistant does not need every email you have ever sent about gardening. It needs the relevant information, including which version is current and which choices are still open. More material can introduce more conflicts if nobody makes those distinctions.

This is also where the caching story reaches its limit. Developers can reduce repeated computation while keeping useful background available. You still need the background to describe the garden you have now. An efficiently processed old plan is an old plan.

Also this week...

OpenAI released GPT-6 Sol and Luna on September 22. The launch announcement positions them as faster, lower-cost alternatives within the GPT-6 family. It lists availability in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, with Luna accessible to Free and Go users in the desktop app. The announcement explicitly says they are not yet available in Chat. Check the surface you are using before assuming a missing model means something is broken.

Anthropic released Claude Opus 5.5 the same day. Its announcement claims about 40 percent lower cost on typical workloads than Opus 5. The listed input and output token prices are 20 percent lower; cache-read prices and the amount of work a model uses also affect the total. That is another reason to read what a price comparison actually measures. These are Anthropic’s results, not a comparison run for this newsletter.

💡 Suggested how-to: trace one detail

Here is an exercise to try with an ordinary project you understand. It has not been run as a controlled test for this issue. Choose a harmless detail whose correct answer you know, such as a garden measurement, the audience for a document, or a preferred date format.

Ask the assistant to use that detail in a small task and identify the source supporting it. You can adapt this request:

Before suggesting the layout, tell me how many garden beds you believe there are and identify the file, message, or instruction supporting that number. If you cannot establish it from the available material, ask me. Treat the attached current plan as authoritative for the dimensions.

Inspect the source it names. If the source is old, correct or replace the relevant material. If the source is current but the answer misstates it, point out the mismatch. If it cannot identify support, provide the missing fact explicitly rather than accepting an invented explanation of where it came from.

Then ask a related question in the place where you expect the information to remain available. Keep the conclusion modest: the new answer shows what happened in that case. It does not establish that every future conversation will retrieve the same detail.

🧰 Tools to try

Your existing memory controls: OpenAI documents these under Settings, Personalization, then Memory, with controls varying by account. Where Sources appear below a personalized response, they may help you inspect what contributed. They are not an exhaustive account of everything affecting the answer.

One current project reference: For details that need exact wording, use a file you can inspect and update. A short, clearly named current plan is more useful for this exercise than another long instruction demanding perfect recall. Keep the original somewhere recoverable if you replace it.

The next time the assistant confidently puts a bench where a shed now stands, ask it to show you the plan it used. You will know whether to correct the plan, the interpretation, or the information available in that conversation.

P.S. An AI wrote these sentences from this week’s primary sources and current product documentation. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.

The following comes from the tweet storm, or threaded essay, pinned to my X account. It was inspired by Naval Ravikant’s “How to Get Rich (without getting lucky)” thread and built around one of my favorite topics and biggest goals: thinking clearly. I consider being called a clear thinker a much better compliment than being called “smart” or “intelligent.” I want to be known as a clear thinker. I hope these 50 maxims and the explanations behind them help you become a clear thinker too. Enjoy.

Clear thinking starts with enough attention to understand a problem and a willingness to test your explanation against reality. AI can suggest options and objections, but you still have to check the important claims and decide what evidence supports them. Before committing to an idea, write down what you expect and try a small test that could change your mind. Use what happens to improve your next decision.

*You can read the essay straight through or return to the maxim that fits a problem you’re facing.

• 01. AI lowers the cost of producing an answer. It does not lower the cost of believing the wrong one.

A polished answer can make you feel as though the difficult thinking has already happened. The sentences connect, the numbers look plausible, and the conclusion sounds reasonable. But none of that tells you whether the assumptions underneath it are true. AI makes it much easier to produce something worth considering. You still have to establish whether it’s worth believing.

Suppose a shop owner asks whether they should open a second location. AI produces a convincing plan based on steady demand, affordable staffing, and customers behaving the same way in a new neighborhood. If any of those assumptions fail, the owner’s actual money is still gone. The cheap part was getting the plan. The expensive part was committing to it.

So use the answer to identify what needs checking. Compare the sales assumptions with your own records, get real quotes, and test demand before signing anything. A quick answer is useful when it accelerates that investigation. It’s dangerous when its confidence lets you skip it.

• 02. Clear thinking means knowing what you believe, why you believe it, and what would change your mind.

You can hold a strong opinion without understanding what holds it up. Sometimes it’s evidence. Sometimes it’s a person you trust, a painful experience, or an idea you’ve repeated so often that it now feels obvious. Clear thinking starts by making that support visible.

Take the belief that remote work makes a team less productive. What does “productive” mean here? Finishing on time? Producing better work? Being available when someone needs help? Once you’ve defined the claim, write down why you believe it and what result would count against it. You could compare similar projects during a trial, including their quality and the effort required to coordinate them.

You don’t need to abandon the belief the moment you encounter one exception. You do need a standard by which it could lose your confidence. If every possible result becomes another reason you were right, you’re protecting a conclusion from examination. Being able to explain what would change your mind is part of having a defensible view.

• 03. Clear thinking depends on a mind that can focus, an explanation that fits reality, and a way to discover when you’re wrong.

You can reason carefully and still reach a poor conclusion because you started with a bad explanation. You can also have a useful explanation and fail to apply it because your attention keeps getting pulled elsewhere. Even when both are working, you need a way to discover whether the decision helped.

Imagine a freelancer trying to earn more. They need enough uninterrupted attention to examine their work. They need an explanation of what limits their income: perhaps pricing, demand, or the hours each project consumes. Then they need feedback that distinguishes those possibilities. Raising prices won’t solve a shortage of qualified inquiries, and attracting more clients won’t help much if every project already loses money.

When thinking goes wrong, inspect the part that failed. Were you too distracted to understand the situation? Did your explanation miss something? Did you act without checking the result? Those problems require different repairs. More concentration cannot make a false assumption true, and reading another book cannot replace feedback from an actual customer.

• 04. Before calling yourself undisciplined, inspect your sleep, your surroundings, and the size of the task. You may be judging your character by conditions you can change.

“I have no discipline” sounds like an explanation, but it’s often just a description of the problem with an insult attached. You didn’t begin the work, so you call yourself someone who doesn’t begin work. That gives you very little to change.

Look at the conditions first. Are you trying to make a difficult decision after a poor night’s sleep? Does every notification interrupt you? Is the task “sort out my life,” which contains about a hundred decisions and no obvious starting point? Any of those can make beginning harder without proving anything permanent about your character.

If you’ve been avoiding a report, try a quiet place and a first step as specific as opening last month’s numbers. Change one condition at a time when you can, so you learn what helps. Responsibility still matters: you’re responsible for improving the conditions you can influence and seeking help with the ones you can’t. But blame is a poor substitute for diagnosis. Give yourself a problem you can actually work on.

• 05. Write complicated problems down. Every detail you must remember competes with the attention you need to examine it.

Try comparing three apartments while keeping every rent, commute, deposit, move-in date, and doubt in your head. Part of your attention will be spent simply preventing those details from disappearing. You’ll revisit the same points because you can’t reliably see what you’ve already considered.

Put them somewhere you can look at. Separate what you know from what you’re assuming, then write the decision you’re actually trying to make. “Which apartment should I choose?” becomes easier to examine when you can see that one exceeds your budget and another isn’t available before your current lease ends. A problem that felt emotionally enormous may be controlled by two practical constraints.

Writing also exposes gaps. “The commute seems fine” becomes “I haven’t tested the commute at the time I’d actually travel.” Now you have a useful next step. The page doesn’t need to be beautiful, and you don’t need a new productivity system. It needs to hold the details still long enough for you to examine how they fit together.

• 06. Expertise is compression. Learn how facts fit together, and you have fewer separate things to hold in mind.

When you understand a subject, you stop treating every detail as a separate instruction. You recognize relationships. Those relationships let you make sense of more information without consciously working through every piece from scratch.

A beginner cook might memorize rules about heat, moisture, and how much food to put in a pan. An experienced cook sees that a crowded pan traps moisture and makes browning harder. That explanation connects several observations at once. It also helps when the recipe changes, because the cook understands what the instructions were trying to accomplish. This is why being able to recite a definition isn’t the same as understanding it. Ask what causes what, when the pattern applies, and what would happen if one condition changed. Then practice with different examples. AI can help generate those examples or explain a relationship, but familiarity with its wording isn’t proof that you’ve learned it. Try using the idea without looking at the explanation. You’ll quickly find the parts you only recognized on the page.

• 07. Find your defaults by asking what happens when you do nothing. Choose that outcome deliberately. It will govern the days when you are tired.

A default is the result your existing setup produces unless you intervene. Your phone opens to something. Your money stays somewhere. Unclaimed time gets filled by whatever is easiest to reach. Doing nothing still leads to an outcome.

Suppose you want to read more, but your phone is beside the couch and your book is upstairs. Reading is possible, but scrolling has the easier starting position. On a tired evening, that difference may decide what you do before you consciously weigh the options. Leaving the book where you sit changes the default without requiring a fresh burst of determination.

Look at what happens on an ordinary bad day, when you’re busy and mildly irritated and haven’t planned everything. Is that setup moving you toward something you want? If it isn’t, change what happens automatically. Defaults deserve occasional review because your circumstances will change. Their value is that you don’t have to rebuild a sensible decision every time you face it.

• 08. Make useful actions easy to begin and distractions inconvenient to reach. A change in friction can outlast a burst of motivation.

Friction is whatever stands between an intention and the first action: a missing file, a login, a cluttered desk, a decision about where to start. None of these needs to be a huge obstacle. A small obstacle repeated at the wrong moment can be enough to send you toward something easier.

Suppose you want to practice guitar after dinner. If the guitar is in its case, in a cupboard, behind three other things, practice begins with a minor household project. Put it on a stand beside the chair, with the next exercise ready, and the first useful action becomes much easier. Put the television remote somewhere less convenient and you’ve changed the competing option too.

You still have to choose to practice. The environment can’t supply a goal you don’t care about. What it can do is stop making the choice unnecessarily difficult. Keep protective friction where mistakes are costly, such as reviewing a message before sending it. Make the useful action easier, not every action indiscriminately faster.

• 09. Do not negotiate with temptation in real time. Move the phone, block the site, or remove the option while you are still thinking clearly.

It’s easy to make a sensible plan when the temptation is somewhere else. The argument changes when the phone lights up or the website is already open. Now the immediate reward is vivid, while the benefit of staying focused is something you’ll receive later.

Make the decision earlier. If you want an hour to write, put the phone outside the room and block the sites you usually visit before you begin. You aren’t relying on winning the same argument thirty times. You’ve changed whether the argument needs to happen at all.

Choose a barrier proportionate to the problem. A temporary website block is different from locking yourself out of something you may urgently need. Leave sensible exceptions and a way to revise a rule that isn’t working. The purpose is to carry a decision you made thoughtfully into a moment when you’re likely to bargain with it. Knowing your future self may be tempted gives you a chance to help them in advance.

• 10. Make your environment work even when your self-control is weak. Removing a distraction reduces how often you need to resist it.

A good working environment should help when you’re having an average day. If it only works when you’re well rested, enthusiastic, and perfectly disciplined, it’s relying on the very conditions it was supposed to support.

Consider someone trying to check a complicated document beside two colleagues having a conversation. Nobody is asking for their attention, and they may never turn around. Yet they keep losing the sentence they were reading and starting again. Telling themselves to concentrate harder doesn’t change the competing speech. Moving to a quieter place might.

That is a different problem from deciding whether to open a distracting app. Some environments keep presenting information you have to process or ignore before you get a choice about pursuing it. The useful intervention may be less noise, fewer visible competing tasks, or a clear agreement about when colleagues can interrupt you.

Try the same kind of work in a quieter setup and compare errors, progress, and how often you need to restart. One good session won’t prove a universal rule about offices, but it can help you identify a condition worth changing. Keep the arrangement simple enough to use on a difficult day. You want your surroundings to absorb some of the effort instead of repeatedly demanding more of it.

• 11. An interruption costs more than the time it lasts. Before switching tasks, write down the next step so you don’t have to reconstruct your thinking.

When you return from an interruption, the document may still be open, but the thought isn’t necessarily where you left it. You have to remember what you were trying to do, which options you ruled out, and why that unfinished sentence mattered. A two-minute interruption can therefore create more than two minutes of lost progress.

Before switching, leave a useful return point. “Continue proposal” tells your future self almost nothing. “Check whether the client’s deadline includes revisions, then finish the delivery paragraph” preserves the decision and the next action. If you stopped because something was unclear, write the question down too.

This won’t make interruptions free, and it doesn’t justify filling your day with them. It makes an unavoidable switch less costly. The note can be a single sentence at the bottom of the document. What matters is that it restores the context you would otherwise have to reconstruct. Write it for the version of you who comes back distracted, not the version who currently remembers everything.

• 12. A mental model simplifies reality by leaving things out. That is what makes it useful, and dangerous: when it disagrees with reality, reality wins.

A map that included every blade of grass would be a terrible map. Its usefulness comes from leaving things out so you can see the roads. Mental models do something similar: they emphasize a few relationships so you can reason about a complicated situation.

The danger appears when an omitted detail controls the result. “Lower prices attract more customers” can be useful, but it leaves out whether people associate the lower price with poor quality, whether you can serve the extra demand, and whether each sale remains worthwhile. A model can be sensible within its limits and misleading outside them. Use it to form an expectation, then compare that expectation with what happens. If the result disagrees, inspect the assumptions and the quality of your observations. You might have measured poorly, or you might need a better model. Either way, the model doesn’t get to overrule the result because it sounds elegant. Its job is to help you navigate what’s actually there.

• 13. A compelling story can leave out as much as any model. Before a narrative decides for you, ask what it left out.

A compelling story usually has a main character, a difficult choice, and a result that makes the choice seem meaningful. Real events contain much more: timing, constraints, other people’s work, failed attempts, and things nobody understood while they were happening. A tidy account selects from that mess.

Imagine reading about a founder who quit their job and built a successful company. The story may accurately describe what happened. It still doesn’t tell you how many people made the same move and failed, what financial support the founder had, or whether customers were already waiting. Those missing details matter if you’re treating the story as advice.

You can enjoy a story and learn from it while asking what it leaves unresolved. Look for the conditions that would have to be similar for the lesson to apply to you. A memorable example shows that something can happen. Establishing how often it happens, why it happens, and whether you should expect it requires more evidence.

• 14. Use AI summaries to decide where to look. Return to the original source before making a decision that depends on what the summary may have omitted.

A summary makes a document shorter by deciding which details deserve to survive. That is useful when you’re trying to find your way around a large amount of material. It becomes risky when the excluded detail is the one your decision depends on.

Suppose AI summarizes a return policy as “returns accepted within thirty days.” The original might limit that promise to unopened products or exclude the item you’re buying. The summary could sound perfectly clear while leaving you with the wrong expectation. Reading the relevant clause is much cheaper than discovering the exception afterward.

Use a summary to locate the passage, identify questions, and decide which sources deserve closer attention. Then inspect the original wherever wording, context, or an exception could change your action. You don’t have to read every page of every source. You do have to check the parts carrying the decision. If the original remains unclear, say so and investigate further. Asking AI to repeat the summary more confidently won’t recover a condition you haven’t checked.

• 15. When reality surprises you, examine the assumption behind your expectation. Investigate the mismatch before explaining it away.

Surprise is useful because it identifies a difference between what you expected and what happened. Something in your picture of the situation was incomplete, or the result included variation you hadn’t allowed for. Either possibility gives you something to investigate.

Suppose you send an offer and receive far fewer replies than expected. “People just don’t get it” protects your original view, but it doesn’t help you improve anything. Check whether the message arrived, whether the recipients had the problem you assumed, and whether the offer was clear enough to evaluate. Also check whether your expectation was reasonable in the first place.

One surprising result doesn’t automatically disprove your explanation. Small samples can mislead, and circumstances can change. The point is to pause before defending yourself. Write down which assumption the result puts under pressure and what information would help resolve it. You want to become quicker at noticing a mismatch, not quicker at producing a reason it doesn’t count.

• 16. Notice what makes you flinch. Sit with it for ten seconds and ask what it’s pointing at.

Sometimes you notice resistance before you can explain it. A question makes you tense. A message stays unopened. A particular line in someone’s feedback makes you want to stop reading. That reaction is worth examining because it may point toward something you’ve been avoiding.

Give it a little space. Ten seconds isn’t a magic threshold; it’s a reminder to pause before distracting yourself. Name what happened as plainly as you can: “I don’t want to check the numbers because they might show this project isn’t working.” Once the concern is visible, you can investigate it rather than letting the discomfort choose for you.

A flinch doesn’t prove the criticism is correct. It can also come from an unfair accusation, a painful association, or a real boundary being crossed. Treat the reaction as a signal to ask a better question. What exactly am I afraid this means, and what evidence supports that interpretation? You can take the feeling seriously without treating it as the final verdict.

• 17. Question every unexamined “should.” Ask who installed it and whether you actually want it before you obey it.

“I should want this” can hide a decision you’ve never actually made. The expectation may have come from your family, your profession, your friends, or a version of yourself whose priorities have changed. Repetition can make it feel like a fact about how life must be lived.

Suppose you believe you should become a manager because that’s what progress looks like in your field. Look at the actual work: hiring, difficult conversations, planning, and being accountable for other people’s results. You may want that. You may prefer becoming excellent at the craft you already practice. The title alone doesn’t settle the question.

Trace the expectation back to its reason. Some “shoulds” reflect real responsibilities, including promises you’ve made and people who depend on you. Others are preferences you’ve inherited without examining the cost. Questioning them doesn’t mean automatically rejecting them. It lets you keep the commitments you can defend and reconsider the ones you’re following mainly because you’ve never stopped to ask.

• 18. Charlie Munger said, “Show me the incentive and I’ll show you the outcome.” Before predicting what people will do, ask what they are being rewarded for.

People respond to more than what a system officially says it values. They also respond to what earns money, approval, promotion, relief, or protection from blame. When those rewards conflict with the stated goal, the rewards deserve your attention.

Imagine a support team paid partly according to how many tickets it closes. The company says it wants customers helped, but the easiest way to improve the rewarded number may be to close difficult cases prematurely. You don’t need to assume the employees are unusually dishonest. You’ve created a situation in which thorough work can make their performance look worse.

Before blaming the people, examine the arrangement. What behavior gets rewarded immediately? What good work goes unnoticed? What happens to someone who reports a problem honestly? Incentives aren’t a complete explanation of human behavior; people also have values, relationships, and professional standards. But if your plan requires people to repeatedly act against what the system rewards, you’ve made the desired result harder to achieve. Change the reward or add a check that makes shortcuts visible.

• 19. A rising metric can hide a worsening result. Measure closed support tickets alongside reopened ones. The number should remain accountable to the thing it represents.

A number is useful because it stands in for something you care about. Response time stands in for service. Views stand in for attention. Completed tasks stand in for progress. The mistake is forgetting that the number and the underlying result can separate.

Suppose a team closes twice as many support tickets this week. That looks good until you discover that many customers are reopening the same unresolved problems. The original metric rewarded an administrative action, and the dashboard mistook it for a successful outcome. Tracking reopened cases makes that difference harder to hide.

Choose a second measure that would worsen if someone improved the first by taking a shortcut. If you’re increasing output, check quality. If you’re cutting delivery time, check errors and rework. You don’t need a dashboard with forty competing targets, each creating more work. You need enough evidence to notice when the number is improving at the expense of the thing it was supposed to represent. Occasionally inspect the actual work too. Two imperfect metrics are still representations.

• 20. Before removing a rule, understand the problem it was built to solve. You can remove the rule and inherit the problem.

An unnecessary rule creates friction. A necessary rule can create friction too. Looking at the inconvenience alone doesn’t tell you which kind you’re dealing with.

Suppose every customer order requires a second person to check the delivery address. It seems wasteful, so you remove the step. Later you discover that the rule followed a series of expensive shipments to old addresses copied from previous orders. The review may still be a poor solution, but the problem it addressed hasn’t disappeared just because the review annoyed you. Ask what happened before the rule existed, which failure it prevents, and whether the relevant conditions have changed. Perhaps the order system now validates addresses reliably. Perhaps a targeted check would replace the blanket review. History gives you something to test, not a permanent veto on change. Once you understand the protection, you can remove it deliberately, replace it, or run a limited trial while watching for the original failure. That’s a stronger improvement than simply making the process shorter.

• 21. Evidence matters when it separates competing explanations. Ask whether you would expect to see the same thing if your belief were false.

Evidence earns its value by helping you choose between explanations. An observation can be consistent with your belief and still tell you very little, because you’d also expect it if your belief were wrong.

Suppose you think a potential customer is eager to buy because they said, “Send me some details.” A serious buyer might say that. So might someone being polite or trying to end the conversation. The reply fits both explanations, so it shouldn’t move your confidence very far. A request to discuss delivery dates with the person who approves the purchase would usually provide more information, though it still wouldn’t guarantee a sale.

Ask what each explanation predicts, then look for the difference. If customers love the product, what should you observe beyond compliments? Repeat use, perhaps, or willingness to pay. Compare that with what you’d expect from people who merely like you. You aren’t demanding perfect proof. You’re looking for information that makes one account of the situation more plausible than its competitors.

• 22. Turn percentages into people. Ten out of a thousand is easier to grasp than one percent.

A percentage can sound dramatic while hiding the scale of what happened. If a website’s sales rise from one to two, sales have increased by 100 percent. If they rise from a thousand to two thousand, the percentage is identical, but the practical result is very different.

Keep the count and the total group together. Ten responses out of a thousand invitations means something different from ten out of twenty. Also check the time window and who was counted. A result drawn from existing customers may tell you little about people who have never heard of the business.

Consider a conversion rate rising from two percent to three percent. In a group of a thousand visitors, that means twenty buyers becoming thirty. The increase is ten buyers. Ten divided by the original twenty is one half, so the relative increase is fifty percent. But the rate itself rose by one percentage point, from two to three.

All of those descriptions are compatible. They answer different questions. The concrete count helps you understand what changed before the most impressive-sounding version decides how you feel about it. Ask for the original count, the new count, and the group they came from whenever a percentage is doing the persuading.

• 23. Begin with what happens in similar cases. Your plan needs evidence for being exceptional, not merely a story about why it is.

Your own plan comes with details that make it feel different. You know your motivation, the care you’ve put into it, and the reasons it could work. Comparable projects usually appear as bare outcomes, without the equally convincing stories their owners told at the beginning.

Suppose you’re planning a paid workshop and expect almost everyone you invite to register. Before building the budget around that expectation, look at comparable workshops: similar audience, price, topic, and relationship with the organizer. What share of invitees actually paid? This starting pattern is often called a base rate. It gives you an expectation grounded in more than the appeal of your own plan.

Choose the comparison carefully. An established teacher inviting past students has a different starting position from a stranger advertising to people who have never heard of them. If you can’t find a close comparison, acknowledge that uncertainty instead of inventing a precise benchmark.

Then adjust for advantages you can defend. Perhaps prospective attendees have already asked for this exact material, or several have made a commitment to attend. You can still believe your workshop will do better. You need to explain why the difference should change the outcome, rather than treating your enthusiasm as the evidence.

• 24. Break a large unknown into smaller estimates. Use ranges. Find the assumption that changes the answer most, and investigate that first.

A large unknown often becomes easier to estimate once you identify what produces it. “How many people might attend this workshop?” is vague. The number you can realistically reach, the share likely to register, and the share of registrants likely to attend give you separate questions you can investigate.

Suppose you can reach about two hundred relevant people. If ten to twenty percent register and around three quarters attend, you might plan for roughly fifteen to thirty attendees. Those figures are hypothetical assumptions, not a forecast you should trust because it contains multiplication. Their purpose is to expose the logic behind the answer.

Now vary the inputs. If a small change in registration rate greatly changes the result, learning about demand may be more useful than refining a minor cost estimate. Watch for assumptions that move together: reaching a larger, less relevant audience might reduce the registration rate. A range is honest only when it reflects plausible conditions. Breaking the problem down makes uncertainty inspectable; it doesn’t make uncertainty disappear.

• 25. When comparing options, consider what you could gain or lose and how likely each outcome is. Write down your estimates so you can see which assumptions your decision depends on.

We often compare options by picturing their most appealing or frightening outcome. A new project offers freedom. A familiar job offers security. Those impressions can hide several possible outcomes with different chances of happening and different costs.

Write down the main possibilities over the same time period. For a small project, that might include strong demand, modest demand, and little demand. Estimate what each would mean after expenses and time, then assign rough probabilities that add up to the whole set of possibilities. You’re making your judgment visible so you can examine it. Decimal places won’t make a guess more trustworthy.

The useful discovery is often which assumption controls the choice. Perhaps the project looks attractive only if customers arrive immediately, while a slow start makes it unaffordable. Now you know what to investigate or protect against. Also include outcomes that don’t fit neatly into money: stress, family time, obligations, and the ability to change course. The calculation should inform the decision you’re actually making, including the parts it cannot precisely measure.

• 26. Weigh both the likelihood of an outcome and its consequences. Then ask whether you can survive being wrong. Favorable odds are little comfort if one loss ends the game.

A favorable average can hide a downside you cannot afford to experience even once. The same uncertain opportunity can be sensible for one person and disastrous for another because they have different obligations and different capacity to recover.

Imagine two small businesses considering the same untested expansion. One can fund a limited trial while continuing normal operations. The other would have to commit the cash it needs to pay staff next month. Even if both owners believe the expansion is likely to work, the second is exposed to a very different failure. A good story about the upside doesn’t make payroll optional. Examine the worst plausible outcome, how it could arise, and what it would leave you able to do next. Reduce the commitment, stage it, or decline it if the downside exceeds what you can responsibly absorb. “Likely to succeed” and “safe enough to attempt” answer different questions. Protecting your ability to continue gives future good decisions somewhere to take effect.

• 27. You do not need certainty to act. You need enough evidence for the size of the bet, a limit on the downside, and a way to learn from the outcome.

Waiting for certainty can feel responsible when it’s really avoiding a decision. At the same time, acting quickly can feel brave when you haven’t understood the downside. The useful standard sits in the relationship between the evidence and the commitment.

You need less evidence to test a new workshop with a small group than to rent a permanent venue for a year. The small trial can reveal whether people attend, whether the material helps, and what takes more work than expected. If it fails, you should still be able to recover and use what you learned.

Define the limits before you begin: the time or money available, what you won’t put at risk, and when you’ll review the result. Make sure the test can actually answer the question controlling the next step. A harmless experiment that teaches you nothing relevant isn’t much help. The aim is to take an action proportionate to what you know, with enough protection that being wrong remains useful rather than overwhelming.

• 28. A good decision can lose. A bad decision can win. Judge the reasoning using what was knowable at the time, then learn from what happened.

After a result arrives, it’s tempting to treat it as a grade on the decision. Success makes the choice look intelligent. Failure makes the warning signs look obvious. But you made the decision before you knew which outcome would occur.

Suppose an outdoor event has a backup venue because the forecast shows a meaningful chance of rain. The day stays dry. That doesn’t automatically mean arranging the backup was foolish. You paid to protect against a possibility that happened not to occur. Conversely, ignoring a serious risk doesn’t become careful planning just because you escaped the consequences once.

Review two things separately. First, did the choice make sense given the information available, the alternatives, and the downside at the time? Second, what does the outcome teach you about those assumptions? A repeated pattern of misses may reveal a broken process even when each individual miss has an excuse. You want to avoid learning the wrong lesson from luck while remaining fully willing to improve the reasoning.

• 29. Write down your prediction and reasoning before the result arrives. Otherwise you can remember being right about something you never predicted.

Once you know what happened, it’s surprisingly easy to remember that you expected it. A vague concern becomes a precise warning in hindsight, and an outcome you barely considered becomes “what I thought would happen.” A written record gives you something less convenient to negotiate with.

Before an important choice, record the outcome you expect, when you expect it, your confidence, and the main reasons. Define the terms. “The trial will go well” is hard to evaluate. A prediction such as “At least eight of the ten participants will complete the exercise without extra help by the end of Friday’s session” gives you a checkable outcome and a deadline.

Then add why you expect that result. You might have tried the instructions with two beginners and found that both finished, while still being uncertain about a larger group. Recording that evidence matters: if the trial fails, you can examine whether the instructions were unclear, the earlier test was unrepresentative, or your confidence simply exceeded what two people could establish.

When the result arrives, compare it with the original words before explaining the difference. Record what you’ll change next time. This doesn’t need to become a diary of every trivial decision. Use it where repeated learning or a meaningful commitment makes an honest comparison worth the effort.

• 30. Estimate from the time similar work actually took. A plan built from perfect days is a promise to be surprised.

A plan often describes the work as though nothing interrupts it: write the draft, review it, make revisions, send it. Real projects contain questions, waiting, errors, missing information, and work that takes longer once you can finally see what it involves.

Look at how long comparable work actually took, from the point you started to the point it was usable. If a previous presentation required research, review, and revisions, those stages belong in the new estimate too. Counting only the hours spent building slides gives you a forecast for an imaginary version of the task.

Use a range when uncertainty is meaningful, and separate your likely finish from a safer commitment. Then adjust for actual differences in scope or experience. You may be faster now, but name what has changed rather than assuming determination will cancel every delay. Historical estimates are useful precisely because they include some of the mess that a fresh plan makes easy to forget.

• 31. Compare your initial time estimates with how long the work actually took. Use the pattern to adjust future estimates for similar tasks.

An estimate becomes more useful when it participates in a feedback loop. Otherwise, you can spend years being “usually a bit optimistic” without learning which tasks you misjudge or by how much.

Keep a simple record of what you expected and what the work required. Be consistent about the measure: active work hours and calendar time answer different questions. A task can need two hours of effort while taking three days to finish because someone else must review it. Both may matter to the next promise you make.

Group similar work and look for a recurring miss. Perhaps you estimate writing reasonably well but forget the time needed to find source material. Perhaps revisions keep doubling a supposedly quick design task. Correct the missing stage before applying a blanket multiplier to everything. Revisit the pattern as your skills, tools, and scope change. The point is to become more accurate, not to turn a small collection of old projects into another rule you’re unwilling to update.

• 32. Every commitment spends hours that cannot be spent elsewhere. Name the best alternative you are giving up before you say yes.

Saying yes spends part of a finite week. The cost includes whatever you can no longer do during that time, even if no money changes hands and nobody sends an invoice for the lost opportunity.

A one-hour meeting may consume the only uninterrupted hour you had to finish a proposal. Add preparation and the effort of returning to the interrupted work, and the commitment is larger than the calendar block suggests. Naming the displaced work makes the trade visible in a way that “I’m busy” doesn’t. Compare the new commitment with your best realistic alternative, not an imaginary hour of perfect productivity. Rest, relationships, and unstructured time can be worthwhile alternatives too. You don’t have to turn every evening into an earnings calculation. You do need to understand what you’re giving up, especially when your calendar fills through small agreeable answers. A commitment can be good in isolation and still be the wrong use of the time you actually have.

• 33. Before delegating, count both the money you will spend and the time needed to explain, check, and fix the work. Ask whether the time you actually free up is worth that cost.

Hiring someone to do a task removes only the work they can complete without you. You may still need to explain the goal, supply materials, answer questions, inspect the result, and request changes. Ignoring that remaining work makes delegation look cheaper than it is.

Suppose a recurring report takes you four hours. Someone else prepares it, but briefing and review take ninety minutes. You’ve freed about two and a half hours, not four, and you still have the fee to consider. If the first handoff also requires building a template, that setup may be worthwhile over many repetitions even if the initial attempt saves little time.

Decide what those freed hours are for and what quality must survive the handoff. Better work, reliable coverage, and less exhaustion can justify delegation as well as additional income. Give the person a clear example of an acceptable result and a way to surface uncertainty early. You want a fair comparison between the full alternatives, not a flattering comparison between their fee and your imagined hourly rate.

• 34. Find the few efforts that produce most of the result. Equal time is a poor default when the returns are unequal.

A full calendar can hide how unevenly your effort contributes to the result. Some conversations lead to most of your useful opportunities. Some product problems cause most of the complaints. Some study habits produce understanding while others mainly produce the feeling of studying.

Look backward at actual outcomes and trace where they came from. If a few types of customer conversation repeatedly reveal important problems, protect time for them before polishing low-impact internal documents. If practice questions expose what you don’t understand, they may deserve more of your study time than rereading familiar pages.

This is a pattern to investigate, not a promise that every situation follows an exact eighty-twenty split. Be careful with work whose benefit is prevention: backups, maintenance, and checking mistakes can look unproductive because success means a problem never appears. Distinguish low visible output from low value. The aim is to allocate effort according to its contribution while preserving the supporting work that allows the most valuable activities to succeed.

Also check whether the apparent winner is repeatable. One large client may account for most of this month’s income, but depending entirely on that client creates a vulnerability. A successful launch may owe part of its result to months of quieter preparation. Concentrate effort where it contributes, while preserving enough support and alternatives that one change cannot remove the whole result.

• 35. When every step must work, one weak step can defeat the whole plan. Improve the bottleneck before polishing what already works.

Some results require a sequence of essential steps. A customer must discover the offer, understand it, decide to buy, complete payment, and receive what was promised. A serious failure at any required stage can prevent the overall result, even when the other stages are excellent.

Suppose a workshop attracts plenty of interested visitors, but the booking form fails on their phones. Improving the headline may attract more people to the same broken form. Fixing the booking problem matters because it lets interest you’ve already earned turn into attendance. This is a weak link: attempts fail as they pass through it.

A queue creates a related but different problem. Every booking might be processed correctly, but if one person must approve each one and cannot keep up, customers wait. This is a capacity bottleneck. The first problem needs a more reliable step; the second may need less work per booking, more capacity, or a simpler approval rule.

Check both failed attempts and waiting time before choosing what to improve. Then measure the final result, not just how busy or efficient one stage looks. Once the limiting point improves, another step may become the constraint. Keep examining the whole path rather than endlessly polishing the part you happen to enjoy working on.

• 36. Getting better at your current method can hide a better method. Test alternatives while the old one still pays for the experiment.

Getting skilled at a method gives you a reason to keep using it. You work faster, make fewer mistakes, and know what to expect. That competence can also make alternatives look worse because you’re comparing your practiced current method with your clumsy first attempt at a new one.

Imagine someone manually assembling the same weekly report. They become extremely fast at copying, sorting, and formatting. A better data export might eliminate much of that work, but learning it would initially slow them down. If they judge only the first attempt, they’ll keep improving a process that may no longer deserve the effort.

Protect the working method while testing an alternative on a limited example. Include setup time, reliability, quality, and the likely number of future repetitions in the comparison. Novelty alone is no reason to change. But neither is your existing competence proof that you’ve found the best route. Give the alternative enough room to demonstrate its value without gambling the output people are already relying on.

• 37. When you have time and results have flatlined, try new approaches. When time is short and results are climbing, double down on what works.

Experimenting spends time in exchange for the possibility of finding something better. Repeating a known approach uses what you’ve already learned to produce a result. Both are useful, but their value changes with your situation.

If you’ve been promoting an event for weeks with little response and still have time, another identical attempt may teach you very little. A different message or audience could reveal something useful. If a particular approach is producing registrations and the event is tomorrow, a major redesign may consume the time you need to do more of what is working.

Check the evidence behind the apparent trend. One unusually good result doesn’t establish a reliable winner, and a short flat period doesn’t prove the method is exhausted. Also consider what failure would cost and how much time remains to recover. You may keep a small experiment running beside the main approach. The useful question is how much learning is still worth buying before you need the result, given what you currently know.

• 38. Treat avoidance as something to investigate. A task may be too vague, too large, or aimed at a goal you no longer want. More pressure will not tell you which.

Avoidance tells you that you’re not doing something. It doesn’t, by itself, explain why. You may not understand the task, may expect it to be unpleasant, may lack a needed skill, or may no longer care about the outcome. More pressure can intensify the discomfort without separating those explanations.

Start with the task itself. “Build my website” contains decisions about the offer, the audience, the words, and the design. “Write one paragraph explaining who this helps” is something you can attempt. If making the next step concrete reduces the resistance, ambiguity was probably part of the problem.

If the resistance remains, investigate further. What do you expect to happen when you start? Is there a missing resource or a conversation you’re avoiding? Do you still endorse the goal? Not every worthwhile task will feel appealing, and discomfort doesn’t automatically mean you should quit. You want to identify whether the useful response is clarification, practice, support, acceptance of temporary discomfort, or a deliberate change of direction.

• 39. Attach each goal to a trigger you cannot miss. “After I put down my coffee, I open the draft” beats “I should write more.”

“I want to write more” leaves the timing undecided. Each day, you have to notice the intention, remember why it matters, and choose a moment to begin. A specific cue removes one of those decisions by connecting the action to something that already happens.

Choose a trigger you can recognize without interpretation, such as putting down your breakfast plate. Attach a small action: open the draft and write the next sentence.

The sequence is specific enough to observe: breakfast finishes, the plate goes down, the document opens, and you write. Keep the document easy to reach so the cue doesn’t lead immediately into a search for the right file.

Then check both parts of the arrangement. Did the trigger happen and did you notice it? If you noticed it but didn’t act, was the action too large or poorly matched to that moment? A breakfast cue won’t help if your mornings are consistently rushed. Adjust the sequence based on what actually happens. The cue gives the goal a place in your day; the small action gives you a practical way to begin when enthusiasm is absent.

• 40. For a small, reversible problem, set a five-minute timer. Finish it, discard it, or define the next physical action. “Think about it later” is not an outcome.

A minor unresolved task can keep returning to your attention without receiving any useful work. You remember it, feel the irritation, and postpone it again. The problem may be small, but the repeated interruption isn’t doing you any favors.

Give it a short, bounded attempt. If you need to find a receipt, spend five minutes looking in the likely places. You may find it, decide it no longer matters, or discover that the next step is requesting a copy from the vendor. Each result changes the state of the problem. “I’ll think about it” leaves you exactly where you began.

The timer limits the investigation; it doesn’t require a reckless decision when it rings. If the issue is larger than expected, define the next physical action and give it an appropriate time. Keep this technique for small, reversible work. Its purpose is to turn a vague open loop into something finished or actionable, not to impose an arbitrary deadline on a choice whose consequences you haven’t understood.

• 41. A cheap experiment can settle what another week of argument cannot. Choose the smallest action whose result could change your mind.

A disagreement can last because both sides keep rearranging the same information. One person thinks customers want a simpler offer. Another thinks they want more features. A longer meeting may produce more arguments without producing anything that distinguishes those beliefs.

Find the smallest test that could change the decision. You might show comparable customers two clearly described offers and measure a meaningful response. Decide beforehand what you’ll count, how long you’ll run the test, and which result would make you continue, change, or stop. Compliments aren’t an adequate substitute for a purchase if willingness to pay is the question.

A cheap test still needs to be informative. A handful of friends may give you useful feedback on clarity while telling you little about demand in a wider market. Keep the conclusion within what the test can establish, and seek stronger evidence as the commitment grows. The point is to spend a little effort making contact with reality when more discussion has stopped changing anyone’s understanding.

• 42. Imagine the plan has failed. Describe how it happened. Look for a plausible cause you can prevent or a warning you can catch early.

Plans are usually described as a sequence in which things go right. A premortem temporarily starts from the opposite end: the plan has failed, and you’re explaining the cause. That change makes it easier to discuss a weakness before anyone has to defend an actual mistake.

Suppose you’re organizing a small event. “Something could go wrong” is too vague to help. “The speaker canceled that morning and nobody had the material needed to replace them” identifies a specific dependency. You can now prepare a backup or decide what you’d tell attendees.

Generate plausible causes, then separate what you can prevent from what you can only detect early. Choose a few safeguards proportionate to the risk. You don’t need to build an elaborate defense against every imaginable disaster, and vivid failure stories aren’t probability estimates. Return to the reasons the plan could work after you’ve checked the risks. A useful premortem improves the attempt; it shouldn’t become a sophisticated way to frighten yourself out of making one.

• 43. Write your own view before asking AI for advice. Otherwise its first answer can become the assumption every later question protects.

The first coherent explanation you read can quietly establish the terms of the discussion. Once AI says the problem is your pricing, your next question may become how to change the pricing. You can spend a long conversation improving an answer to a question you never independently chose.

Before asking, write a short view of your own: what you think is happening, what evidence you have, what remains uncertain, and which decision you need to make. You might discover that you don’t yet have a view. Record that honestly, including the facts you would need to form one.

Now ask AI to examine the situation and compare its explanation with yours. Where do they differ, and what would settle the disagreement? The first note gives you a record of your starting point and makes changes in your reasoning visible. It doesn’t make your initial view superior. You should change it when better evidence warrants it. You’re creating a chance to think before a fluent answer makes one particular framing feel inevitable.

• 44. A question that contains your preferred answer invites agreement. Give AI the problem, the evidence, and the constraints before telling it what you hope is true.

If you ask, “Why is my plan such a good idea?” you’ve already assigned the model its conclusion. Even a softer prompt such as “Help me explain why this is the right move” directs attention toward supporting arguments. The response may be useful advocacy, but it shouldn’t surprise you when it agrees.

For an assessment, give the situation first. Describe the options, relevant evidence, constraints, and what success would mean. Ask what each option gets right, where it could fail, and what information is missing. Research on sycophancy in language models has documented assistants favoring agreement with user beliefs, which is one reason to avoid treating a pleasing response as independent confirmation.

Your preferences still belong in the decision. Wanting more family time or less travel is relevant information, not an error to conceal. The distinction is between giving the model your actual values and feeding it the conclusion you want endorsed. A neutral prompt helps you examine the trade. It doesn’t guarantee an unbiased or correct answer.

• 45. Ask AI for the strongest objection to your plan and the strongest case for the alternative. Give it a disagreement worth investigating.

A useful objection identifies something that could make your plan fail or make another option better. “There are risks” is too vague. “This requires ten hours of work each week, and you currently have three available” gives you a concrete conflict to examine.

Ask AI for the strongest supported objection, not a long list of everything that could conceivably go wrong. Also ask it to make the best fair case for the alternative. That second request can reveal benefits you’ve minimized because you already prefer your own plan, such as reliability, simplicity, or the freedom to change direction later.

Then investigate the claim doing the real work. Are ten hours actually required? Could you reduce the scope? Does the alternative really avoid that constraint, or merely hide it somewhere else? Don’t manufacture equal strength between options when the evidence favors one. The purpose of the exercise is to find an objection that deserves your attention and determine whether it changes the choice. Producing disagreement is only the beginning.

• 46. An AI criticism needs checking as much as an AI compliment. A model can invent reasons you are wrong just as fluently as reasons you are right.

Criticism can sound more sophisticated than praise. A model that finds a flaw appears to be thinking independently, especially when it uses confident language and explains the objection neatly. But it can produce an unsupported criticism with the same fluency as an unsupported compliment.

Suppose AI says your article contradicts itself. Ask it to identify the two passages and explain the conflict. You may discover a real inconsistency. You may also discover that one passage applies to small reversible decisions while the other applies to commitments with serious consequences. The criticism missed a distinction the article already made.

Turn the feedback into claims you can inspect. Check quotations against the text, factual assertions against sources, and predictions against whatever evidence is available. Revise when the criticism survives that check. Reject it when it doesn’t. Asking another model can surface a different view, but agreement between models isn’t automatically independent confirmation. What makes feedback useful is the reasoning and evidence underneath it, not whether the tone is flattering or severe.

• 47. Count the cost of explaining, supervising, checking, and repairing delegated work. An answer generated in seconds can still take hours to trust.

The time shown on an AI generation is only one part of the job. A report may appear in thirty seconds and still require extensive checking before anyone should rely on it. If you count the generation time while ignoring the verification, you can mistake transferred work for eliminated work.

Consider a spreadsheet summary. You need to provide the correct data, explain which rows count, check that the totals match, and investigate suspicious conclusions. If you already have a dependable formula that produces the answer, a long AI conversation may be the slower route. For a different task, AI might remove substantial work even after review.

Compare the complete process with a realistic alternative. Define an acceptable result before delegating, and make errors easy to detect. Start with bounded work whose inputs and outputs you can inspect. As reliability improves, the appropriate scope may grow. The test is whether the whole assignment becomes easier to complete to the required standard, including the cost of fixing mistakes and the consequences of missing them.

• 48. For a consequential decision, state what you chose and why in your own words. An explanation you cannot give without the model is a decision you are not ready to own.

You don’t need to personally know every fact behind a decision. We all depend on other people’s expertise. You do need enough understanding to explain the choice you’re making, the advice you’re relying on, and the risk you’re accepting.

Close the chat and describe the decision to someone who hasn’t read it. What are you choosing? What evidence supports it? What is the strongest reason against it? What would cause you to reconsider? If your explanation reduces to “AI said this was best,” you haven’t yet made the reasoning your own. Notice where you get stuck. Perhaps you understand the recommendation but not the probability attached to it. Perhaps a key term remains vague. Ask for an explanation, inspect the source, or consult someone qualified for the decision. Being able to repeat fluent language isn’t the acceptance check. You should understand enough to identify the assumptions, recognize an important exception, and explain why the choice fits your actual circumstances.

• 49. Let being wrong cost you an opinion before it costs you another year.

An opinion becomes expensive when protecting it determines what you continue to do. You keep pursuing an offer people don’t want, defending a method that no longer works, or waiting for a situation to become what you originally expected. The cost grows while the belief stays comfortably familiar.

Suppose you believe a certain service is what your customers need. Repeated conversations point to a different problem, but you’ve already built the materials and told people your plan. Changing direction now would be embarrassing. Continuing may feel easier today while consuming months you could use to build something more useful. Separate the work you’ve already spent from the value of the next commitment. What would you choose now if you encountered the same evidence without needing to defend your earlier prediction? That question doesn’t require abandoning a worthwhile effort after every setback. It requires giving new evidence a fair chance to change your course. You can acknowledge that an opinion was reasonable when you formed it and still decide it no longer deserves another year.

• 50. Borrow ideas. Take responsibility for what you do with them.

You should learn from other people. Their ideas can spare you mistakes, reveal options you hadn’t considered, and give you a clearer way to understand what you’re experiencing. Taking responsibility doesn’t require pretending you invented everything yourself.

It does require judging where the advice applies. A method that works for an experienced business owner with savings and an established audience may be a poor fit for someone starting from scratch. An AI explanation may be useful while containing one assumption that makes its recommendation wrong for you. Credit the source, inspect the relevant evidence, and adapt the idea to the conditions you actually face.

That applies to this essay too. Pick a maxim that speaks to a real problem, explain why you think it fits, and try an appropriate action. Then pay attention to what happens. If the result challenges the advice, investigate the difference. Borrowing an idea gives you something to work with. What you choose to do next, and how honestly you learn from it, remains your responsibility.

By Payton Bilodeau

P.S. If you are still confused on any of these, take the time to sit with them and think about what they mean, and/or use an LLM to help you understand them.

If you use ChatGPT Plus or Pro, there is a small setting worth checking before your next difficult request. OpenAI changed how the product decides when to spend more time thinking, and adding “think harder” to your prompt will not change that setting for you.

That is easy to miss because the conversation still looks like a conversation. You can ask for a dinner suggestion, a cleaner paragraph, or a project plan full of competing deadlines in the same box. The request becomes more demanding without the interface necessarily making the change you expected.

The useful response is to get more deliberate about where you need extra reasoning. Some work deserves it. Some work needs a missing fact, a clearer requirement, or a simple calculation instead. Learning to tell those cases apart will help more than adding increasingly stern instructions to every prompt.

📌 TL;DR

• Check the control → ChatGPT Plus and Pro no longer automatically move from Instant to deeper thinking for difficult requests. Choose an available thinking level when the work calls for it.

• Diagnose the failure → Extra reasoning may help connect the facts you supplied. It cannot establish a fact you left out, and it should not quietly change a requirement to make a plan work.

• Check the result → A quick, natural reply tells you little about whether the underlying work is correct or complete. This becomes especially important as voice systems talk while doing other work.

Change the setting when the facts have to fit together

In its September 14 release notes, OpenAI says it is retiring automatic switching from Instant to Thinking for Plus and Pro users globally. Those users can still select a deeper thinking option themselves. The change concerns how the product routes a request; it does not establish that the underlying models became less capable.

OpenAI’s current thinking guide makes the practical detail explicit: asking ChatGPT to “think harder” or “think deeply” does not automatically select a different thinking level. On Plus, Medium and High are available. The available choices differ by plan, and managed workspaces have their own controls, so check what your account actually shows.

A request to be careful can still improve the instructions. “Check whether any dates conflict” tells the assistant what to examine. But the words in the prompt and the selected amount of thinking are separate controls. You can ask a useful question while leaving the product on Instant.

There is a good reason to keep a fast option. If you need three alternatives to an awkward sentence, the quality is usually easy to judge as you read. You can reject the wording immediately, ask for another version, and get on with the document. A long wait would need to buy something you value.

Compare that with asking whether an exhibition can be ready before the doors open. The framed pictures arrive at a particular time. Hanging takes a fixed amount of time, photographs must happen afterward, and the opening is already advertised. Every individual sentence can be clear while the combined plan is impossible.

The second task requires the assistant to preserve relationships between facts. Moving photography earlier is only a solution if the pictures are already on the wall. Adding another volunteer only helps if the hanging time can change. A tidy timetable can conceal either mistake, which makes a first glance a poor test.

That is a sensible trigger for trying an explicit thinking level: the answer depends on several conditions being true together. You might be comparing proposals with different exclusions, checking a schedule, or finding the consequences of changing one step in a process. Choose the effort because of the work the answer must do.

Even then, the setting is an opportunity for a better answer, not a certificate attached to it. If a response is wrong, you still need to identify the mistake. Otherwise, you can spend longer waiting for a more elaborate version of the same misunderstanding.

Find the mistake before you ask for more thinking

Take a deliberately simple, fictional exhibition plan. The framed pictures arrive at 10 a.m. Hanging them takes 90 minutes and cannot begin before delivery. Photographing the finished display takes another 30 minutes and must happen after all the hanging. Both jobs must be finished before the doors open at 11:30.

Those requirements cannot all hold. Hanging finishes at 11:30, and photography finishes at noon. The plan is half an hour short even before allowing for a late van or a missing screw. That conclusion follows from the supplied times; no knowledge of the art world is required.

Suppose an assistant returns a polished schedule with photography beginning at 11. The problem is specific: it has overlapped two jobs you said must happen in sequence. Asking for a more enthusiastic plan will not help. Asking it to check each start time against the previous task’s finish gives it a concrete error to repair.

If that kind of missed dependency keeps happening, trying a higher thinking level is reasonable. You have a known failure to test. Keep the same facts and requirements, then inspect whether the new answer finds the conflict without changing them.

Now remove the delivery time from the request. You still know that hanging takes 90 minutes and photography takes 30, but the assistant no longer has enough information to say whether the morning works. A careful answer should identify the missing arrival time or give a condition: delivery must happen by 9:30 to finish at 11:30, with no spare time.

Extra thought cannot tell it when the van is booked. It might propose an earlier delivery, but that is an option to investigate. It becomes a fact only when somebody confirms it. This is why “use a smarter model” can be an expensive answer to an information problem.

There is another version of the mistake. Perhaps the assistant notices the clash and changes hanging from 90 minutes to 60. The arithmetic now works, but it has solved a different problem. If 90 minutes was a firm requirement, the response should say which condition needs to change and leave that decision with you.

A useful revision would be: “The current plan finishes at noon. To keep the 11:30 opening, we would need delivery by 9:30, a confirmed shorter installation, or permission to photograph afterward.” You can then check which option is possible. The assistant has made the decision clearer without pretending it has already been made.

Notice how the right next step changes across these examples. A missed connection calls for checking the reasoning. A missing arrival time calls for information. An impossible combination calls for changing a real constraint, with the relevant person’s agreement. All three can produce an answer that looks wrong, but they should not produce the same follow-up request.

You do not need to label every conversation this carefully. Use the distinction when you catch yourself asking for a third rewrite and the same underlying problem survives. Before sending another prompt, name one thing the answer failed to preserve. That often reveals whether the assistant needs more effort, better information, or a decision only you can make.

A smooth conversation still needs a correct result

The interface is becoming less helpful as a clue to how much work is happening underneath it. Google’s Gemini 3.8 Live announcement, published September 15 and updated September 17, describes a version called Extended Thinking that can reason while speaking. Google also says these models can carry out tool calls in the background while the conversation continues.

That is Google’s description, not a hands-on result from this newsletter. But the design has an immediate implication: a spoken acknowledgment can arrive before the task is finished. An assistant saying it is checking the delivery booking may still be waiting for the booking system.

The reverse inference is weak too. A long pause does not show that a schedule was checked correctly. You need the resulting times, the conflict it found, or the confirmed booking. How the conversation feels and what the assistant accomplished deserve separate attention.

There is a useful research example behind this distinction. Google cites EVA-Bench, a ServiceNow research framework published in May. Its researchers evaluated 12 voice-agent systems using simulated conversations across 213 scenarios in airline service, healthcare HR, and IT support. They measured accuracy separately from conversational experience.

For task completion, the framework checks whether the simulated system’s final records match the expected result. Other measures examine whether the agent follows the supplied information and speaks details correctly. The experience measures cover things such as conversational progress and turn-taking. A system can therefore be evaluated for making the right change as well as for being pleasant to talk to.

This was a benchmark using simulated callers, not a study of people doing their everyday work. Its older model results should not be treated as a verdict on this week’s Gemini release. The useful lesson is the measurement design: checking the conversation alone leaves part of the job unchecked.

Back at the fictional exhibition, “I’ve moved the delivery to 9:30” would need a different check from “9:30 is the latest arrival that makes the schedule fit.” The second is a calculation you can inspect. The first claims that an external booking changed, so you would want to see the updated confirmation before telling the installation team to arrive earlier.

This distinction also helps you decide whether a slower mode earned its time. Did it find an overlap you missed? Did it ask for the fact needed to answer? Did the confirmed action match what you authorized? Those are useful improvements. A longer explanation with the same wrong arrival time is more reading.

OpenAI’s September 16 guide to connecting AI usage with business value makes a related point for organizations: examine the kind of work being done and include review and correction time when evaluating the benefit. You can borrow that question without needing an enterprise dashboard. Judge whether the whole task became easier to finish correctly.

Also this week...

ChatGPT is now available inside Word. OpenAI’s September 17 update adds drafting, summaries, selected-text revisions, and formatting through the Microsoft add-in. It is available across ChatGPT plans, including Free, subject to usage limits. If you already work in Word, that may remove some copying between windows; inspect the changed passage in the document before accepting it.

Anthropic is bringing evaluators inside the lab. Its September 18 partnership announcement says Faculty, Accenture’s specialist AI business, will lead work on model evaluations, safeguards, and alignment with internal access. Anthropic will fund the work directly, and reporting and access standards are still being developed. More access could make scrutiny better informed, but this is an announced arrangement, not a set of completed independent findings.

💡 Suggested how-to: repair one disappointing answer

Here is an exercise you can use with a low-stakes task you understand well. It has not been run as a controlled comparison for this issue. Its purpose is to help you find the right correction before changing several things at once.

Start with an answer you rejected and write down the exact defect. “It was bad” is too broad to guide the next attempt. “It scheduled photography before the display was finished” gives you a check you can repeat.

Compare the answer with your original request. If you never supplied the relevant fact, add it. If the fact was present but the answer ignored its relationship to another fact, keep the information unchanged and try an available deeper thinking setting. If the requirements contradict one another, ask for the conflict and the smallest change that would resolve it.

You can adapt this request:

Check this plan against the requirements below. Treat the confirmed times and durations as fixed. Identify any task that starts before its prerequisite finishes. If the plan is impossible, show the conflict and the earliest possible finish. Keep proposed changes separate from confirmed facts, and ask for missing information instead of inventing it.

Then check the output against the defect you named. For the fictional gallery schedule, the answer must recognize noon as the earliest finish and the 30-minute shortfall. A more elegant table does not compensate for missing that conflict.

Keep the result in proportion. One successful attempt is a reason to try that approach again on similar work, not proof that a setting is universally better. If the same failure remains, stop cycling through increasingly elaborate prompts and inspect the underlying facts or work through that step yourself.

🧰 Tools to try

ChatGPT’s thinking control: Use the plan-specific guide to find the options your account supports. Try a deeper setting on a task with interacting constraints, while keeping the request stable enough to judge the difference. You do not need to upgrade a subscription merely to begin checking your current setup.

A plain calendar or spreadsheet: Put each task’s start, duration, and prerequisite where you can see them together. For a short schedule, that may reveal the conflict faster than another conversation. AI can help construct the plan while the visible rows give you an ordinary way to inspect it.

The next time an answer disappoints you, point to the first thing it got wrong. If the missing piece is a fact, supply it. If the facts were there and the relationship was missed, check the thinking setting before trying again. You will have given the next answer a specific job to do.

P.S. An AI wrote these sentences from this week’s primary sources. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.

You know where you want the picture to go. You can see the empty space for the headline. Then you open an AI chat and spend five minutes trying to describe a layout you could draw in ten seconds.

By the third correction, the prompt has become a small essay. “A little further left” now refers to something that moved two versions ago, and the part you liked has somehow changed too. You started with a simple idea and ended up explaining your explanations.

This week’s AI releases offer a reason to reconsider that habit. OpenAI updated image creation, voice, and access to source files. None of these updates invented the idea of showing or telling a computer what you want, but together they invite a useful question: which part of this request is getting lost when you turn it into typed instructions?

Sometimes a drawing preserves the relationship between things. Sometimes talking helps you discover a constraint you hadn’t thought to include. Sometimes the answer is already in a document, and your summary leaves out the detail that matters.

The opportunity is to spend less effort translating what you know, while staying precise about what AI should do with it.

⚡ TL;DR

• A rough sketch can communicate a layout before you have the words for it. Use written instructions for the details the drawing cannot settle.

• Voice can help you work through an unfinished thought. Confirm the resulting decisions in writing, especially names, numbers, and dates.

• Give AI the actual source when you want an explanation grounded in records. A connected file still needs a clear question and the right interpretation.

• Try a second input method on one small task this week. Keep it only if it reduces the total work, including corrections.

A rough sketch can carry what a paragraph loses

Imagine you’re making a poster for a neighborhood plant swap. You want a large plant on the right, the event name on the left, and a quiet strip along the bottom for the time and location. You have a clear arrangement in mind, even if you couldn’t name a single design principle behind it.

A drawing with three rough shapes can express that arrangement. It doesn’t have to resemble a finished poster. Its job is to show which element occupies which space, how large the main object should feel, and where nothing should compete for attention.

OpenAI’s September 8 ChatGPT Images 2.5 announcement claims improvements in following visual references, making focused edits, and maintaining consistency across revisions. Those are vendor claims, rather than results from a test for this issue. They matter here because a reference loses much of its value if the system keeps wandering away from it.

The practical starting point is more modest than asking for a perfect design. Give the system your arrangement and ask it to develop the visual treatment around that arrangement. A useful instruction for the imaginary poster might be: “Use this sketch for placement. Keep the plant on the right and leave the bottom strip clear. Make the illustration warm and simple.”

That separates two decisions you might otherwise blur together. You have chosen the layout, while the system has room to suggest the finish. If the result feels wrong, you can identify whether it ignored your arrangement or made a stylistic choice you dislike.

There’s now a built-in Sketch option in the ChatGPT mobile app. Type @ in the message box, choose Sketch, draw, confirm it, and add written instructions before sending. Image creation is available across tiers, although particular features and allowances differ.

Drawing also has limits. An empty rectangle could mean “put the headline here,” “leave this empty,” or “this is a separate panel.” Label the few parts whose meaning could be mistaken, then explain what the model may change. The drawing communicates relationships; the accompanying words remove ambiguity.

For our poster, the event name, date, and venue still deserve exact written instructions and a careful check in the finished image. A beautiful plant cannot rescue the wrong address. If you’re still deciding the date, leave that area empty instead of letting a plausible date become part of the design by accident.

The same principle applies when revising an image. Point to the part you mean and say what should change, but inspect the rest afterward. OpenAI’s image documentation warns that an edit can extend beyond the selected area, so a local request is no guarantee of a local result.

This is useful when you keep correcting position, proportion, or empty space. If all you need is a plain sentence on a plain background, typing may already be the simplest route. The reason to draw is that the task contains a relationship that your paragraph is struggling to carry.

A conversation can uncover the request you haven’t formed yet

Now imagine you’re organizing the plant swap itself. You ask for an event plan, but halfway through explaining it you remember that several attendees will be beginners. Then you realize that helping them choose a suitable plant matters more than fitting in as many activities as possible.

You could type all of that. But if the useful details tend to emerge while you talk, a conversation gives you somewhere to work them out. You can reconsider an idea, respond to a question, and hear whether the explanation makes sense before turning it into a brief.

The September 9 ChatGPT Voice update adds GPT-5.6 or GPT-6 Astra for searching and harder reasoning, with availability depending on your plan. Separately, OpenAI’s September 10 GPT-Live-1 API announcement describes a voice system that can listen and speak simultaneously. That second announcement is aimed at developers building products, so it shouldn’t be read as a promise that every feature is present in every ChatGPT voice conversation.

For an ordinary planning session, the useful move is to ask for questions before a plan. You might say: “I’m working out a beginner-friendly plant swap. Help me identify what I haven’t decided. Ask one question at a time, and don’t turn a possibility into a commitment.”

Suppose the conversation brings up a plant-care table. You mention that a friend might staff it, then spend several minutes exploring how it could work. By the end, it may feel settled simply because you’ve talked about it in detail, even though your friend hasn’t agreed.

That’s where a written closing step earns its place. Ask for a short account of what you decided, what remains tentative, and what needs a fact checked. Read it yourself and correct any upgrade from “could” to “will” before using it to organize the event.

There are also two different voice jobs worth separating. The Voice guide distinguishes conversation from Dictation, which produces text you can edit before sending. If you know what you want to say and just dislike typing on a phone, Dictation may be enough.

For important details, the editable text is valuable. You can inspect the spelling of a person’s name or replace “next Saturday” with the exact date. The same guide says Voice transcripts aren’t verbatim records, so don’t treat a fluent conversation as a perfect set of minutes.

One limitation matters if you plan to combine this with the next section: Live currently cannot find or add files from Library, and it does not initially support connected apps or plugins. Use voice to explore the problem, then move to a supported file-based chat to check the records. Those are separate capabilities, even when they appear inside the same product.

You haven’t gained anything if you spend twenty minutes talking around a question that takes one sentence to ask. Voice earns its time when the conversation helps you notice something that belongs in the request. Once the request is clear, you can stop exploring and use it.

Give AI the source before asking for the explanation

A week before the event, you want to know whether registrations are low. You tell AI, “Only twelve people have signed up. How can I improve attendance?” It can now suggest better promotion, more reminders, or a different message, all while knowing almost nothing about whether twelve is a problem.

Perhaps the room holds fifteen people. Perhaps last year’s event also had twelve registrations at this point and finished full. Or perhaps this year’s sheet counts households while last year’s counts individual attendees. Your short summary can be accurate and still leave out the comparison needed to answer your question.

Source files give the system a chance to inspect those details. In its September 10 release notes, OpenAI says Box, Dropbox, and SharePoint are joining Google Drive in ChatGPT Library. Eligible web accounts can browse connected material and add selected files or folders to a conversation, with source citations and existing permissions applying. Mobile support is due later.

For the hypothetical event, you could select this year’s registration sheet and last year’s record. Then ask the system to compare registrations at the same number of days before each event, identify what each row represents, and show the entries behind its answer. That is a more answerable question than asking why attendance is bad before establishing whether it is.

The distinction between a count and an explanation matters. A sheet might show fewer registrations without showing why they fell. If it contains no evidence about promotion, weather, or competing events, the system should present those as possibilities to investigate, rather than explanations discovered in the records.

You can make that boundary explicit: “Use these files to establish what changed. Tell me which comparisons are valid and which information is missing. Keep possible causes separate from findings supported by the files.” This gives the analysis a useful stopping point when the documents cannot answer the whole question.

For people working with larger organizational datasets, OpenAI also introduced the Data plugin for ChatGPT Work and Codex this week. It’s intended to connect business questions to data and relevant context. Its setup documentation makes clear that Data and the appropriate sources need to be installed or connected, with authorization where required.

That is a different starting point from attaching two sheets for a small event. A workplace may already have agreed definitions for customers, completed orders, or revenue. Those definitions belong in the analysis because a correct calculation of the wrong thing can produce a convincing but useless answer.

You don’t need to connect a whole folder to benefit from source-based work. Start with the smallest set of relevant files you’re allowed to use. Open the cited source for a consequential finding and check that it actually supports the claim, including its date and the conditions around it.

A source link makes that check possible; it doesn’t perform the check for you. The gain is that you can return to the evidence instead of arguing with a summary built from your own incomplete recollection.

🛠️ Suggested how-to: compare two ways of asking

Here’s a small experiment based on the documentation above. It hasn’t been run as a comparative test for this issue, and it doesn’t require you to change your whole workflow. Choose a low-stakes task where you already know enough to judge the answer.

1. Try your usual route first. Make a small poster concept, work through an event decision, or compare two simple records. Keep the request and note roughly how long you spend explaining and correcting it.

2. Change the input on a second attempt. Use a rough sketch for the layout, a short conversation for the undecided problem, or the actual files for the comparison. Keep the desired result the same so you can tell whether the input helped.

3. Check the details that matter. Did the layout survive? Did a tentative idea become a commitment? Did the comparison use the same dates and units? Count those corrections as part of the work.

4. Keep the route that costs you less effort overall. Include drawing, talking, finding files, reviewing, and fixing. A faster first answer can still leave you with more work at the end.

One attempt won’t establish which method is generally best. It can tell you whether a particular recurring frustration has an easier route. If the sketch takes longer than the explanation, or the conversation sends you into more indecision, you have a useful reason to keep typing for that task.

🧰 Tools to try

ChatGPT Sketch: Start here if you know how something should be arranged but keep struggling to describe its position. The mobile instructions are the quickest way into that experiment.

ChatGPT Dictation: Choose this when the request is already formed and you want to review the words before sending them. The Voice guide explains the distinction from live conversation.

Data in Work or Codex: Consider this if your workplace already has a supported, authorized data source and a real question for it. Read the setup requirements before assuming it’s ready to use.

For your next request, notice what takes the most effort to communicate. If it’s the arrangement, draw it. If it’s the thought you’re still forming, talk it through. If it’s what happened, bring the records. Then use exact words for the details you need AI to preserve.

P.S. An AI wrote these sentences from this week’s primary sources. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.

Think of a useful thing you’ve been meaning to make: a guide for a new colleague, a comparison of the software your team uses, or a proper explanation of a process that currently lives in eight documents and somebody’s head. You could probably ask AI for help with each piece. You would still have to turn the pieces into something another person could use.

That is the kind of job worth trying again after this week’s model releases. OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.1 both put more emphasis on sustained work. The interesting experiment is to give one of them responsibility for a complete, bounded project, then find out how much work is actually left for you.

There is a good version of this, where the neglected guide finally gets made. There is also a version where you acquire forty pages of homework. The difference deserves more attention than another benchmark chart.

📌 TL;DR

• Start with the finished object. Describe what someone should be able to use at the end, including the awkward cases a summary would leave out.

• Give the work somewhere to remember itself. Longer projects need an explicit record of sources, decisions, and unfinished work.

• Make the result easier to check. Ask for the supporting evidence alongside the deliverable, and count your review time when deciding whether the experiment worked.

Start with the finished object

In its Astra announcement, OpenAI describes stronger multistep work and better adherence to document, spreadsheet, and presentation templates. The company is rolling it out to Plus, Pro, Business, and Enterprise subscribers within existing usage allowances. Access can arrive gradually, and Enterprise administrators need to enable it.

Anthropic’s Fable 5.1 release makes a similar case for longer assignments. Its current product page lists access for Pro, Max, Team, and Enterprise users. These are the companies’ descriptions, not a comparison I’ve run on your work. They justify a fresh trial if you have access; they don’t establish that you should buy another subscription.

Suppose you organize community workshops. Volunteers answer the same questions about arrival times, room setup, cancellations, and equipment. The answers are scattered across eight approved documents. You’ve wanted a simple volunteer guide for months, but pulling it together keeps losing to the next workshop.

An obvious request would be to summarize the eight files. That might produce eight accurate summaries while leaving the original problem almost untouched. A new volunteer still has to work out which instruction applies, find the relevant exception, and connect the answer to the event they’re helping with.

A more useful assignment is to produce the guide itself. Tell the assistant who will read it, what that person needs to do, which documents are authoritative, and where the finished file should live. Include a sample of the writing or layout you want. Ask it to bring conflicting instructions back to you instead of silently choosing whichever sentence sounds more plausible.

The distinction shows up in the output. A summary might say that the venue has rules for equipment. A usable guide tells the volunteer where to collect the projector, what to check before the room opens, and whom to contact when the cable is missing, provided the source material actually contains those answers.

You are giving the assistant a more demanding job because the pieces have to fit together. It must notice when two sections use different names for the same room, avoid repeating the cancellation policy in three places, and leave a clear question when the source doesn’t explain something. Those connections are much of the work you were postponing.

Start by asking for one representative section. Choose something with a little friction, such as opening the room, where several documents might overlap. If the sample merely rearranges sentences, correct the assignment before asking for the whole guide. If it answers the questions a volunteer would actually have, you have a reason to continue.

This small sample also gives you a fairer test of the new model. A vague request can make a strong model look unhelpful, while an easy request can make a weak process look finished. One section with a real complication gives both of you something specific to improve. The method was useful before these releases; a stronger model gives you a reason to retry a job that previously needed too much help.

Give the work somewhere to remember itself

Once the assistant starts combining documents, its job includes remembering decisions. Perhaps you confirm that the newer venue instructions replace the older ones, or decide that the guide should cover volunteers rather than paid instructors. Those answers need to survive the next ten changes.

OpenAI describes an experimental Astra feature in Codex that keeps notes across context windows and lets the model search earlier windows. A context window is the amount of material a model can work with at once. This feature is intended to preserve details during long sessions; it isn’t yet a universal default.

There is an instructive failure in Anthropic’s new Fermat formalization report. Early attempts lost track of project state. The researchers describe progress after adopting Prove2Me, which recorded the relationships between mathematical tasks and helped agents find and reuse completed work. The organization surrounding the model contributed to the result.

Your volunteer guide doesn’t need a mathematical research platform. It does need somewhere to record that the September venue sheet supersedes the June sheet. A short project note can carry that decision, the reason for it, the remaining questions, and the location of the latest draft.

For example, imagine one source says volunteers should arrive at eight and another says eight thirty. The assistant asks, and you confirm that eight applies to instructors while volunteers arrive later. Saving only “arrival time: eight thirty” would lose an important distinction. Saving the two roles and their respective times protects the next section about who opens the building.

That is why a useful progress note records decisions rather than a diary of effort. “Read the files and worked on the guide” tells the next session almost nothing. “Arrival times differ by role; the organizer confirmed the distinction; equipment instructions remain unresolved” gives the work a place to continue.

Keep this record small enough to read. You don’t need a dashboard for a document. You need the current draft, the approved inputs, and a short explanation of the decisions that would otherwise be easy to lose. Ask the assistant to update the explanation when you correct it, and check that important corrections reached the actual guide as well.

This becomes particularly useful when the job resumes later. Instead of reconstructing the conversation, you can ask the assistant to read the project note, inspect the current draft, and continue with the unresolved items. If it cannot access those files, attach them again. A reassuring claim that it remembers is less useful than seeing the right decision in the right section.

When the venue changes its opening time next month, the source record also tells you which sections might need updating. Keep that explanation connected to the guide, and the first version becomes easier to maintain.

Make the result easier to check

The scale of Anthropic’s mathematical result is striking. Its September 4 report describes an eleven-day effort to formalize an existing proof of Fermat’s Last Theorem in Lean, a language whose checker verifies the logical steps. This used an internal research model roughly comparable to Fable 5.1 and about six billion output tokens. It wasn’t an ordinary chat session, and Fermat’s theorem had already been proved.

The report says Lean checked the resulting formalization. An office document has no equivalent test that can establish every part of its usefulness. That limits the analogy: we can borrow the idea of making checks part of the process without pretending that a volunteer guide can be certified like a mathematical proof.

Nathaniel Whittemore makes a related point in AI Daily Brief’s discussion of agentic loops: knowledge work needs deliberately designed ways to decide whether a result is good enough. That is useful advice when an assistant can keep revising long after a normal conversation would have ended.

For the guide, ask for a separate source table. Each instruction should point to the document and section supporting it. Keep the table out of the volunteer-facing copy if it makes the guide harder to read. It exists so the person reviewing the work can quickly compare an instruction with its source.

A link alone isn’t proof. Open a few of the cited sections and check whether they support what the guide actually says. If the guide tells volunteers they may move a projector between rooms, a citation to a document that merely lists the projectors does not establish permission to move them. That small difference could disappear inside otherwise fluent prose.

Then try the guide with questions someone might ask during a workshop. What happens if the room is locked? Where does a late volunteer go? Who can approve a cancellation? A useful draft either provides an answer grounded in the documents or identifies the missing information clearly enough for you to fill it in.

Include at least one question the sources cannot answer. That tests whether the assistant can leave an honest gap. A tidy document with a visible “confirm with the organizer” note may be closer to useful than a seamless document that invented the organizer’s policy.

Some checks can be mechanical, such as confirming that every source appears in the source table or that every section has a heading. Others require judgment. A guide can pass all the mechanical checks and still bury the arrival instructions beneath two pages of background. Read it in the order a volunteer would need it, with their first task in mind.

Finally, count the cost of review. If you spend an hour fixing a document that would have taken you forty minutes to write, that particular attempt hasn’t saved time. It may still have taught you something, but you should be willing to reduce the scope, improve the inputs, or do the remaining work yourself. More revisions are worthwhile only while they are resolving identifiable problems.

Also this week

Anthropic announced Enterprise Frontier Safeguards, designed to let customers keep monitoring data in cloud infrastructure they control and route human review to their own people. The phased rollout is planned for later this fall, with an interim zero-retention arrangement for eligible customers.

This is relevant when a proposed AI project depends on internal documents. It is an enterprise arrangement with eligibility conditions, not a privacy setting every subscriber has today. For the exercise below, use documents you’re already allowed to put into your chosen tool. A model upgrade doesn’t change that permission.

Try this: finish one neglected document

Here is a suggested exercise, not a workflow I’ve personally tested. Choose a small document you’ve been postponing and gather the approved source material. A guide for volunteers works, but so does a checklist for running a meeting or an explanation of how your team names and stores its files.

Pick something you understand well enough to review. Before opening the chat, write one ordinary question the finished document must answer. That gives you a reason to reject a polished result that doesn’t help.

You can adapt this brief:

Create a practical guide for [reader] so they can [specific task] using the attached approved documents. Produce an editable draft in [format]. Use [reference] as the model for tone and layout.

Begin with one representative section about [task]. Flag conflicting instructions and missing facts. Ask me only for decisions that would change the content; record my answers with the source decisions.

After I review the sample, complete the guide. Keep a short project note with the current draft location, settled decisions, and unresolved questions. Provide a separate source table linking each instruction to its supporting document and section.

Check the finished guide against [three realistic questions]. Report anything those checks cannot establish. Make at most two revision passes, then return the draft and any unresolved issues. Leave the original files unchanged and don’t distribute the guide.

When the sample arrives, review the substance before the formatting. Fix a missing exception or confusing instruction first. Once the assistant understands that standard, let it complete the rest and inspect whether it applied your correction consistently.

Keep the best brief and the decisions you made while reviewing. If the experiment works, you’ll have a useful starting point for the next update. If it doesn’t, you’ll have a specific account of what failed, such as incomplete sources or too much rewriting, instead of a general feeling that AI was disappointing.

For the final check, give the guide to someone who wasn’t involved in creating it and ask them to find the answer to that ordinary question you wrote down. Notice where they hesitate. If they can find the answer and act on it without calling you, the document is beginning to do the job you wanted it to do.

P.S. An AI wrote these sentences from this week’s primary sources. Payton set the editorial standards and what good looks like. He wants to be transparent when AI wrote the issue.

That's an extraordinary tool. It also leaves me to decide which argument I should believe and what I'm going to do about it.

Getting an answer and reaching an understanding are becoming dangerously easy to confuse. For me, clearer thinking means taking back what gets my attention, what earns my belief, and what I actually do next.

1. Stop making everything a test of character

Suppose I sit down to write with my phone beside me, my inbox open, and a social feed one click away. Then I get distracted and decide I need more discipline.

I've also asked a half-finished sentence to win a popularity contest against everything on the internet.

Before I judge my character, I want to inspect the setup.

Every environment makes some actions easier than others. A feed that keeps loading gives me no obvious place to stop. A document buried inside six folders gives me several places to give up. Neither arrangement forces my decision, but both influence it, especially when I'm tired.

The useful question is: where does this setup take me when I'm not paying much attention?

If opening my laptop puts yesterday's entertainment in front of me, I've given tomorrow's distraction a head start. I can close those tabs tonight and leave the document open at the sentence I need to finish. I can put my phone somewhere that requires getting out of the chair to retrieve it.

Apparently, becoming a more serious person sometimes involves walking down a hallway.

One decision made calmly can spare me from repeatedly making it while tempted. I still have work to do, but I've stopped arranging the room so that avoiding it is the easiest thing available.

2. Let reality interrupt a good explanation

Once I can stay with a thought, I need to decide whether it deserves my confidence.

A map leaves things out so it's small enough to use. My explanations work the same way: they compress a situation into something I can understand and act on.

The trouble starts when I forget how much I've left out.

If someone hasn't answered my message, I can construct an entire story about how they feel about me. The actual evidence is a message without a reply. The resentment, disrespect, and impending collapse of the relationship are additions from my extremely enthusiastic writing department.

The same mistake can look impressively professional. A plan can be beautifully organized around a customer need nobody has checked. A calendar can be full of exercise sessions that never happen. The description looks healthy while the thing it describes is struggling.

AI makes producing those descriptions much easier. I can turn an untested assumption into a persuasive document before I notice it's still an untested assumption.

So I'd rather use AI to find the assumption doing the most work in my argument, then ask what evidence would support or undermine it. After that, I need to check something outside the conversation. Talk to the potential customer. Look at what I actually completed. Ask the person instead of prosecuting them in my head.

An AI criticizing my idea hasn't proved it wrong, either. Agreement and criticism both give me something to investigate. I still have to do the investigating.

3. Give the thing you're avoiding five honest minutes

Checking reality sounds sensible until reality is sitting in an email I don't want to open.

I can understand every principle above and still spend the afternoon researching a problem because the next actual step might tell me something I don't want to hear.

That urge to look away deserves attention. I'd like to identify the exact part I'm avoiding. Am I missing information? Am I expecting criticism? Do I want the outcome, or have I inherited another obligation I've never examined?

Those are different problems. Calling all of them procrastination makes them harder to solve.

For an ordinary task, I like a modest test: spend five minutes attempting the first concrete step. Open the email. Find the missing document. Write the question I need answered.

The aim is to discover what's actually involved. I might finish. I might discover the task no longer matters. Or I might learn that I need one person's answer before anything else can happen. Even then, I've replaced a vague burden with a specific dependency.

Five minutes won't settle a major life decision. It can help me stop treating an unopened message like a mythical creature.

Before I ask for another explanation, I want to know what the next one will help me do. If I already have enough information to take a small, sensible step, another answer can wait.

The map can stay on the desk for a moment. I'd like to see what's outside.

This is issue 1 of AI Mentorship. If you can open ChatGPT and Notion, you can use what is in here. This is about understanding AI, using it well, and staying optimistic without pretending the hard parts are not there.

This week is a product week, not a speech about the newsletter. The labs did not ship a new flagship model for you to chase. They shipped memory, a trigger, and a browser.

The assistant can carry what you already told it, start when Slack or Gmail moves, and click through pages that never got a connector (a connector is a proper hook into an app, like Gmail or Slack, so the model does not have to click around a website as you). That is more useful than last week’s chat box, and it is also easier to leak a rant into an email, or to click a button you did not mean.

Both things can be true. Pick one control. Edit it. Keep a human on the consequential click.

📌 TL;DR

• Claude memory → Chat and Cowork now share one memory you can read and edit. The catch is that junk and private rants get remembered too, unless you clean them.

• ChatGPT Work triggers → Plus and Pro can start a task when Gmail, Slack, or GitHub moves. Free cannot. Narrow the filter or it will fire on noise.

• Claude’s browsers → Paid Claude can click in Chrome, or use a built-in Cowork browser that is not your tabs. Prompt injection (hidden instructions on a page that try to hijack the agent) is still real.

• Builder’s notes → A 20-minute setup: audit Claude’s memory topics, then write one Notion skill. This is a how-to, not a fake diary.

Claude finally remembers across chat and Cowork

The most annoying part of using two Claude surfaces was briefing the second one like a new hire. You would think something through in chat, then open Cowork (Anthropic’s hand-off mode, where Claude runs a longer task in the cloud instead of chatting back and forth) and type the same facts again.

On August 25, Anthropic made the memory in chat the same as the memory in Cowork. TechCrunch called it the difference between two products under one roof and one continuous assistant, and that is the right read.

Claude now writes topics into memory as you chat, instead of waiting to summarize after you leave. Mention that a deadline moved to September, and the next conversation already has it. You do not have to say “remember this.”

The part that actually makes this safe enough to use is the list. Settings, then Memory, then Topics: every saved topic is a short file you can read, edit, or delete. Fix the company’s old name once, and every later chat and Cowork run gets it right.

Memory is on by default on Free, Pro, and Max, on web, desktop, and mobile. On Team and Enterprise, an admin has to allow it, and it stays off for you until you turn it on. Sensitive topics (health, beliefs, politics, and similar) stay out unless you opt in, and government IDs and a few other categories stay out even then.

Ben Tossell wrote on August 27 that he is not sure shared memory is a good thing. You do not want a frustration you mentioned in a random chat to change the tone of an email Cowork writes later, and he is not wrong.

If the list is editable, the job is to EDIT the list. Open Topics this week, delete anything you would not want sitting next to a client email, and write one sentence about how you like updates, who they go to, and what “done” means on the project in front of you. That is the whole upgrade: a notebook you can actually correct.

Anthropic Claude’s memory works everywhere, August 25, 2026

TechCrunch Claude Cowork finally remembers what you told the app in chat, August 25, 2026

Ben’s Bites Who let the agents in, August 27, 2026

ChatGPT Work can start when Slack or Gmail moves

Scheduled tasks used to mean a clock: every Monday at 9, every Friday at 4. That is fine for a recap. It is useless for the email that actually matters, which arrives at 2:14.

On August 25, OpenAI’s ChatGPT release notes added webhook-triggered tasks in ChatGPT Work. A webhook is a signal from another app that something happened, and the first supported signals are a new Gmail message, a new Slack channel message, and GitHub pull request activity.

Plus and Pro can create these in Work on web, iOS, and Android. Free and Go cannot. Free users can still run up to three ordinary scheduled tasks, once a day at most. If you are on Enterprise or Edu, an admin has to turn on “Allow event-triggered scheduled tasks” first, and it is off by default.

The setup is short. Settings, then Apps, then connect Gmail, Slack, or GitHub. Open Work, describe the event and what ChatGPT should do, and fill the three boxes you will see: Trigger, Condition, Prompt.

For Slack, add @ChatGPT to every channel the task watches. Reactions, edits, deletes, and DMs do not fire it. Only new messages, and that last line is the constraint that saves you.

If your team marks “this needs a human” with an emoji, this feature will not see it. If you watch a busy channel with no sender filter, you will get a pile of summaries nobody asked for.

OpenAI’s own example is the right size. Summarize an email that says registration opened, or draft next steps after a named client leaves feedback in Slack. Use one sender, one channel, and one output you will actually read.

Actions that need approval pause until you review them. Leave that on for anything that sends, files, or spends.

This is the week ChatGPT stopped waiting for you to remember to check. It still needs you to decide what is worth checking.

OpenAI ChatGPT release notes, August 25, 2026

OpenAI Help Scheduled tasks in ChatGPT

Claude can click the pages that have no connector

A lot of real work still lives on a site with no Gmail-style connector: a vendor portal, an internal dashboard, a form that only exists as a webpage.

On August 26, Anthropic made Claude in Chrome generally available on every paid plan. Claude can view the page you are on, click, type, move between tabs, and fill forms, using the logins you already have. It can also approve some actions on its own now, instead of tapping you for every click. A safety classifier (a second check that asks “does this match what the user asked?”) reviews the action before it runs, and you can turn auto-approve off.

The same day, Anthropic shipped a built-in browser inside Cowork on the desktop app, with nothing to install. When a task needs the web, a browser opens in the side panel. That browser is Claude’s, not yours. It does not see your tabs, bookmarks, or passwords unless you bring a login over, site by site, and banking, email, and single sign-on stay out unless you choose to include them.

Use the split the way Anthropic described it. Claude in Chrome is for the page you already have open, with the accounts you are already in. The built-in browser is for handing off research, invoices, or a portal while you keep working.

If you already use the Chrome extension, it stays the default. Otherwise the built-in one takes over once it reaches you.

Pro, Max, and Team are getting it this week, and Enterprise admins can turn it on now. Chrome only, not other Chromium browsers, and not mobile. You still need the desktop app for files on your computer.

The hard part is prompt injection. Hidden instructions in a page, an email, or a form field can try to redirect the agent. Anthropic spent a year on this before calling Chrome generally available.

They say probes scan tool results, and the classifier blocks actions that do not match your request. They also say the risk is reduced, not gone, and you should start on sites you trust.

Take that at face value. Do not point either browser at a bank, a payroll admin, or a page you would not let a new intern click unsupervised. Watch the first run, and if it drifts, stop it.

A browser that works is a gift. A browser you do not watch is a liability.

Anthropic Claude in Chrome is generally available, August 26, 2026

Anthropic Claude gets its own browser in Cowork, August 26, 2026

Anthropic Help Get started with Claude in Chrome

Also this week...

• Slack Code → As of August 26, a coding agent can spin up a dedicated Slack channel for a longer job, with diffs, HTML previews, and a place for the team to review before anything ships. Live today with Claude, Devin, Copilot, and Vercel. ChatGPT is coming soon. You still need access to the partner agent.

• Gemini Live → August 26. Talk, then let it act. Daily Brief (Google AI Plus and up) reads Gmail and Calendar out loud. Spark (Google AI Pro and up) can turn a voice brain dump into a Doc and keep running when you leave. Connect apps in Personal Intelligence first.

• ChatGPT Work, signed-in sites → Also August 25. Work’s cloud browser can pause on a login, let you enter credentials in a form the model cannot see, then keep going. Plus and Pro. Confirm before payments. Still do not start with a bank.

• OpenAI and Cursor → August 28. OpenAI says it will wind down model access in Cursor, proposed shutoff November 12, 2026, after SpaceX bought Cursor. If you live in Cursor, that is a calendar date, not a vibe. Have a second model ready inside the same tool before November.

💡 Builder’s notes

A 20-minute setup, not a diary

Here is a setup you can actually finish before lunch, using things that shipped in the last seven days.

1. Audit Claude’s memory (8 minutes)

Open Claude, then go to Settings, Memory, Topics, and read the files. Delete anything you would not want in a client email. If a topic is close but wrong, edit the sentence so it is true.

Add one short topic of your own, in plain language:

I send weekly updates in five bullets: what shipped, what slipped, what I need, risks, and the next decision. No fluff. No hedging.

That sentence will now sit next to Cowork when you ask for the update. If you are on Team or Enterprise, memory may be off until you or an admin turns it on.

2. Write one Notion skill (12 minutes)

A skill is a reusable instruction stored as a page. Notion’s own rule is the right one: if you would paste the same prompt twice, it is a skill.

Go to Library in the sidebar, open the Skills tab, and add a skill. If you want Notion Agent to pick it automatically, put it in a skills database and write a clear Description. Automatic use only works when that description exists.

Paste this as a starting page, then change the names:

Weekly recap

Description: Use when someone asks for a weekly update, a status recap, or “what happened this week.”

Goal: a sendable update in five bullets.

Inputs: the page or notes the user has selected. If nothing is selected, ask for the project name.

Format:

1. Shipped

2. Slipped

3. Need from others

4. Risks

5. Decision needed this week

Constraints: no adjectives that do not change a decision. No “hope to.” Name owners. If a fact is missing, write UNKNOWN instead of guessing.

Run it once on last week’s notes, then tighten the page when the output is wrong. The skill gets better because you edited the page, not because you wrote a longer prompt in chat.

If you also live in ChatGPT Work, the cousin of this skill is one webhook task for a single sender, not a whole Slack workspace. Same idea: narrow the input, name the output, and a human still sends it.

🧰 Tools to try

• Claude in Chrome → Official Anthropic extension. Paid plans only. Confirm the publisher is Anthropic before you install. (Chrome, not Edge or Arc)

• Notion Skills → Reusable instructions as pages. Library, then Skills. Put them in a skills database if you want automatic use.

• Scheduled tasks in ChatGPT → Clock tasks for Free. Event-triggered Gmail, Slack, and GitHub tasks for Plus, Pro, and eligible workspaces.

• Claude Academy → Anthropic’s free learning site. Start with Claude.ai or Cowork if you are not a developer. Browse without signing in. Sign in to save progress.

🥣 Brain food

• 5 Rules for Better AI Writing → Nathaniel Whittemore (NLW) on The AI Daily Brief, August 27. The purity fight is over. Quality is not. Lock the thesis before you hand the middle to a model, then strip the obvious AI-isms.

• What We Learned From 15 Hours of Anthropic Certification Training → Natalia Quintero at Every, August 31. The courses give a shared vocabulary, not a playbook. Her team’s consensus: Anthropic’s docs beat the videos if you actually want to use the tools.

• Who let the agents in → Ben Tossell, August 27. The memory caveat in beat 1 lives here, plus a clean explanation of ChatGPT Work’s sign-in form. Read him when you want the “I am not sure that’s a good thing” line before you turn a feature on.

If you only do one thing after this letter, do the memory audit, then write the skill. If you have never named how you like work done, that part takes some thinking. The setup itself is simple. Don’t overcomplicate it.

That’s it. That’s the whole thing.

P.S. This issue was written by AI. Payton built the automation, then put taste and judgment on what it was allowed to argue, what it had to cite, and what it was forbidden to invent. Value is still value. When he writes one himself, it will say so and he will sign it.