01 There’s a limit to how much better you can make your request—and you’ve reached it

We call it request the text you type when you want something from AI. In the early years, getting this text just right made all the difference: a ten-line request delivered much more than a one-line request.

Today’s models have changed the game. They understand a short, direct request well. What used to require a whole page of instructions can now be handled with three clear sentences. In other words, beyond a certain point, polishing the text no longer gets you better results.

Helena is a lawyer who handles contracts. She asked for a plain-language summary of a commercial lease to explain to her client. She rewrote the prompt five times—more detailed, more polite, more specific. The summary remained accurate and remained useless: it talked about “the parties” and “the subject matter,” while her client wanted to know when they could leave without a penalty.

What was missing wasn’t a better word. It was for the AI to know that the end reader is a bakery owner, not another lawyer. That information didn’t belong in the request — it belonged somewhere else.

02 What’s missing is almost never the sentence — it’s what the AI doesn’t know

The second layer is the context: everything the AI would need to know about your situation that isn’t in the request. Who the reader is. How your company communicates. What was already decided. Which document is authoritative.

Rafael is the marketing manager for a chain of three stores. He asks for campaign texts and gets phrases that could be from any store in the country. His request is impeccable. What’s missing is for the AI to know that the chain never puts a price in the first sentence, that its audience is local, and that “flash sale” was prohibited by legal last year.

03 Knowing isn’t the same as doing

There’s a third kind of failure, and it’s the most frustrating because the AI seems to have gotten it right. It understood the request, had the knowledge, wrote well—and the work still wasn’t finished. Because no one checked, no one opened it to look, no one compared it with what was already there, no one told it where to stop.

Helena again: the AI summarized the contract accurately, in the right tone, and cited a 90-day termination notice period that wasn’t anywhere in the document. She sent it to the client. The error wasn’t in the writing or the knowledge—it was the absence of a simple check between “the AI finished” and “the client received it.”

This third layer has a name: it’s the work structure around AI. Tools for it to take action, rules for where it must stop, and a required check before anything goes out the door. That’s what almost the rest of this course is about.

04 The three layers work together — they never replace one another

It’s tempting to hear “asking well isn’t enough anymore” and conclude that the request has become unnecessary. It hasn’t. Without a clear request, the best knowledge in the world doesn’t know what task to serve, and the best-built structure executes the wrong thing precisely.

The order is additive, not a swap: the request sets the direction, the knowledge gives AI what it needs to know, and the structure defines how to execute, check, and stop. Each layer only works when the previous one is in place.

For Rafael, this means he won’t stop being careful when making requests—he’ll keep making them well, but from now on with the network’s rules included and a mandatory final review before any text goes live.

01 A request isn’t a sentence — it’s five decisions

When you write “summarize this,” you’ve made one decision and left four open. AI will fill in all four on its own, in the most likely way—and the most likely way is rarely your way.

The five decisions are always the same, for every request: role (what perspective it speaks from), task (what to do, using the right verb), format (the shape of the answer), constraints (what it must not do) and ready (how you'll recognize that the answer works). Only the second one is usually written down.

Helena wrote: “summarize this contract.” Four pieces were missing. With them: “you’re explaining this to a client with no legal background (role); summarize this lease, highlighting deadlines, penalties, and how to end it (task); in five short bullet points (format); without using legal terms unless you explain them (restriction); the response is good if my client understands it without needing to ask me anything afterward (done).”

02 Role and task: who’s speaking, and what verb you used

The role isn’t make-believe. Saying "you’re a copywriter" doesn’t summon a copywriter — but it puts the response in a certain register of vocabulary, rhythm, and priorities. It’s the difference between asking your accountant and your business partner for an opinion: both are competent, but they look at different things.

The task depends on the verb, and using the wrong verb is the silent cause of half the off-target deliverables. “Summarize” makes it shorter. “Explain” expands and translates. “Rewrite” keeps the length and changes the tone. “Compare” puts things side by side. Many people write “improve” — which isn’t a task verb, it’s a wish.

Rafael used to ask, “improve this caption” and get longer captions back. He changed it to “rewrite this caption, keeping it the same length, with the benefit in the first line” and stopped fighting the result. Same tool, same day, different verb.

03 Format and restrictions: the mold and the guardrails

Format is the mold for the response: five bullet points, a three-column table, two paragraphs, an email with a subject line. When you don’t provide the mold, the AI uses its own—usually a medium-length block of text, which is almost never what you’ll paste into your work.

A restriction is the opposite of a template: instead of stating the form, you draw a boundary around what cannot appear. “Don’t make up numbers,” “don’t name competitors,” “don’t promise a delivery date.” Restrictions matter more than adjectives—“be professional” says nothing; “no exclamation points and no slang” says everything.

Sônia is an accountant and sends the same deadline reminder to dozens of clients. The restriction that changed everything for her was one line: “don’t cite any amount or date that I haven’t written here.” Before that, AI filled in plausible dates—and plausible, in a tax notice, is worse than missing.

04 Describe the right answer before you see it

The fifth part is the least used and the one that pays off most: write down, before sending, how you’ll recognize that the answer works. This isn’t praise (“I want a good answer”); it’s a test you can apply yourself in ten seconds.

This changes two things. First, it forces you to know what you want—half of bad prompts come from a request the author couldn’t define in the first place. Second, it gives AI a verifiable target, and AI tends to adjust to targets.

Rafael now writes: “the caption works if a customer can tell, just by reading it, which store it is and what they get by coming in today.” He tests this by reading it aloud. When it fails, he doesn’t rewrite the whole request—he fixes only the part that failed, because now he knows what it is.

01 It knows nothing about your situation — and doesn’t tell you

A chat AI is like a highly trained professional who started at your company today, and no one introduced them to anyone. They read quickly, write well, and know the subject in general. They don’t know your client, your history, or why that word is forbidden here.

The painful difference from a human colleague is that the colleague asks questions. AI doesn’t ask: it fills in the gaps. When information is missing, it chooses the world’s most common version and writes with the same certainty.

Rafael asked for a text about the new store and got a polished paragraph about “personalized service and a welcoming atmosphere”—a phrase that could describe any store in any city. It wasn’t a lie; it was filling in the void. The void was his chain.

02 Four types of knowledge, and only one of them is a document

When people talk about “providing context,” almost everyone thinks of attaching a document. A document is just one of the four types, and not always the most decisive.

The four: facts (the material being worked on — the contract, the numbers, the text); who reads (the person on the other side and what they already know); your way (two or three examples of what you consider good, which teach more than ten adjectives); and what’s already been decided (the rules and prohibitions that came from past conversations).

Helena had the contract (the facts) and nothing else. When she added “the reader is a business owner with no legal training” and pasted an old summary the client had praised (her style), the quality jumped—without changing tools or writing a longer prompt.

03 The dossier: one page you write once and paste in every time

It doesn’t make sense to rewrite the four types every time. What works is having a single page — we call it dossier — with what almost never changes in your work, saved in a regular notes app, ready to paste at the start of the conversation.

A useful dossier fits on one screen and has four short sections: who I am and what I do; who usually reads what I write; three standing rules; two examples that show my tone. None of this is technology secrets—it’s what you’d tell a new coworker over your first coffee together.

Sônia, an accountant, has hers in nine lines: a mixed client base of individuals and businesses, clients who aren’t tax experts, never promise a deadline without checking the official calendar, and two past notices that worked well. She pastes these nine lines before any request and stops explaining the same thing every week.

04 Knowledge gets old — and old knowledge is worse than none

The dossier has a trap: it silently becomes outdated. The store changed its name, the legal rule was dropped, the client changed. And the dossier keeps pasting in the old version with all the authority of something written down.

Outdated information is worse than missing information. When it’s missing, AI writes something generic and you notice. When it’s wrong, AI writes something specific and convincing—and nobody notices.

Rafael started doing something that takes two minutes: on the first business day of the month, he reads through his entire dossier and crosses out what’s no longer valid. That’s how the store’s old name was removed—after it had already been published once.

01 The model is the engine; the structure is the whole car

An engine by itself, sitting on the workshop floor, has power but won’t get anyone anywhere. It needs steering, brakes, a dashboard, a seat belt, and a fuel tank. None of that makes the engine more powerful—it makes that power useful and safe.

It’s the same with AI. The model is the engine: it understands, writes, reasons. The work structure is everything built around it so the work comes out finished, reliable, and within your rules. In technical writing, you’ll see the word harness — that’s exactly it, and it has nothing to do with whether the model is better or worse.

Helena discovered this by comparing two law firms. The other one uses the same AI she does and delivers better contract summaries—not because it has a different engine, but because it has a folder of approved templates, a required reviewer, and a rule that nothing goes out without citing the source clause.

02 Eight pieces—and you already have two

The structure is always built from the same eight pieces: instructions permanent (the rules that always apply), knowledge (lesson 3), capabilities (what it can do beyond writing), specializations (fixed recipes for tasks that repeat), memory (what carries over between conversations), limits (where it stops and asks), review (the step that separates "finished" from "ready") and approval (the human who approves before the thing goes out).

You’ve already built two of these eight in previous lessons. This isn’t a list to memorize; it’s a list to look at a task of yours and ask, piece by piece, “does this exist in my case, or do I do it by hand every time?”

Rafael did this inventory for “publishing a campaign text.” He discovered he had knowledge and instructions, no memory, no written limits, and a check that happened only when he remembered. Three out of eight, in a process he already considered organized.

03 The same task, with and without structure

It helps to see the difference in a concrete task, not in the abstract. “Create a page for my company,” without any structure, goes like this: you ask, the AI generates something, you look at it, don’t like it, and ask again. The work ends when you get tired.

With structure, the same task becomes a sequence: it reads the brand guide, looks at previous pages, uses the standard page template, builds it, opens it in the browser, compares it with the reference, tests the buttons, checks it on a phone, fixes what broke, publishes a test version, and only then asks your permission to go live.

Notice what changed: none of the steps on this list involve model intelligence. They’re all organizational steps—and that’s the kind of organization Helena would call a procedure, and that any serious office has been applying to interns for decades.

04 It doesn’t come ready-made — and it shouldn’t

Nobody builds all eight pieces on a Saturday afternoon. The structure grows through incidents: whenever something goes wrong, you add the piece that would have prevented it. That’s how it grows to match your level of risk instead of turning into bureaucracy nobody follows.

The usual path is: start with a good request, turn it into a request + dossier, add a fixed recipe when the task repeats, add memory when you get tired of repeating yourself, add limits after the first scare, and require a review after the first error that makes it out into the world.

Helena is on this path. She added the rule “never cite a deadline without identifying the clause” afterward—and only after—the summary with the invented deadline. The piece came from the incident, which is why she follows it.

01 Answering is one thing; doing is another

For a long time, AI only returned text. You asked, it answered, and everything else — opening the document, looking up the data, checking the date, pasting it in the right place — was still up to you.

What changed is that it can now take action: read a document you attach, look up current information, examine an image, calculate things from a table, open a page, and check whether it’s live. We call these things capabilities.

Helena spent months pasting contract excerpts into the conversation one at a time because she thought that was how it worked. The day she attached the entire document and asked, “find all the termination clauses and tell me which page each one is on,” she saved forty minutes—and got the page reference, which was what she needed to trust the result.

02 What you can ask for today — and what you still can’t

This part changes over time, so it’s stamped with the date. Today, with a typical chat AI, you can attach documents and spreadsheets, ask it to search for current information, send a photo of a paper document and ask it to read it, ask it to calculate things from a table, and ask it to open a public page to check what’s written there.

What it still can’t do on its own in a regular conversation: access your systems with a password, send messages in your name, or use your computer. These capabilities exist in specific tools, with authorization — and that’s exactly what lesson 7 will cover.

Rafael discovered two capabilities he’d been using poorly. He used to type the campaign numbers into the conversation; now he attaches the spreadsheet. And he’d ask, “check if the page is live,” thinking the AI knew—it didn’t, but when he explicitly asked it to open the address and read the title, it did.

03 Specialization is the house recipe, written once

One specialization is the fixed recipe for a recurring task. It isn’t a better request: it’s the request, plus the steps, plus what to consult, plus what never to do, plus the completion criteria—all together, written once and reused the same way every time.

A good comparison is a sewing pattern. No one drafts the pattern again for every piece; you change the fabric and measurements, and the pattern ensures the piece has the same fit. Specialization does this for your task: it ensures the fifth text of the month has the same quality as the first, even on Friday night.

She has one for “summarize a contract for a client with no legal background”: read the attached document, list deadlines, penalties, and ways to end it, write five bullet points without unexplained legal terms, cite the source clause for every statement, and mark in red what she needs to confirm. She pastes it in and changes only the contract.

04 How to write yours — and how to know it’s good

A good specialization fits on half a page and answers five questions in order. More than that becomes a document no one reads; less than that turns back into a loose request with a different name.

The quality test is simple and ruthless: give the recipe to someone on your team who has never done the task. If they can run it without asking you anything, it’s good. If they need to ask, what they asked is exactly the line that’s missing.

Sônia wrote hers for a tax deadline notice and tested it with the intern. The first question was, “and what about individual taxpayers?”—the missing line, and the same information that had already ruined a notice before.

01 What carries over from one conversation to the next

Within the same conversation, the AI keeps track of everything that was said: you correct it once, and it maintains the correction until the end. When you start a new conversation, that thread is cut. It goes back to being the newcomer from lesson 3.

Today’s chat tools already offer a space to save permanent information — usually called memory or "custom instructions" in the settings menu. There’s little space, and you decide what goes in. That decision is the whole lesson.

Rafael thought the AI had “memorized” the chain’s tone because it always got it right on Thursdays. It hadn’t: he was the one working in the same conversation all week. On Monday, with a new conversation, the tone went back to generic—and he blamed the tool for a bad day.

02 Three things are worth remembering; everything else is for the moment

What’s worth keeping fits into three categories. Who you are and who you work for — your role, your audience, what you produce. The rules that don’t change — the prohibitions, the format you always prefer, the word your organization doesn’t use. What’s already been decided and cost a lot — the correction you made three times, the mistake that went out once.

Everything else is for the moment: today’s document, this afternoon’s question, this quarter’s number. It lives in the conversation and dies with it, and that’s fine—too much memory gets in the way just as much as too little.

Helena saved three lines: that she’s a contract attorney, that her writing is for clients with no legal background, and that no statement about a contract goes out without citing its source clause. She never saves the specific contract for each case—and shouldn’t.

03 What must never go in

Memory is for convenience, not a vault. Never include passwords, information that identifies a client by name, ID numbers, specific contract amounts, or anything you wouldn’t write on a note left on your desk.

The reasoning is the same as what you already use with paper: whatever is kept is exposed to anyone who has access to it. And most importantly, none of this is necessary to get the quality right. The tone, format, and house rules don’t depend on the client’s name.

Sônia, an accountant, solved this with a change of wording: instead of saving “the client Marcos is an individual,” she saved “my client base is mixed, half individuals.” Just as useful to AI, with no name stored.

04 The memory sheet — and the review that keeps it from going stale

In practice, you keep two places and don’t confuse them: the memory worksheet (short, permanent, inside the tool) and the lesson 3 dossier (longer, pasted in when the subject calls for it). The worksheet is what you’d tell anyone working with you; the dossier is what they need for a specific task.

And the same warning from lesson 3 applies, even more strongly now: what’s saved gets old silently, and what grows outdated in memory influences everything, not just one conversation. That’s why the record has a date and a scheduled review.

Rafael reviews his sheet on the same day he reviews the dossier—two minutes, the first Monday of the month. That’s how “anniversary campaign” was removed, after it had been pulling texts toward a date that had already passed for two months.

01 Unlimited autonomy is the surprise that hasn’t happened yet

While AI only wrote, the worst-case scenario was bad text. You read it, didn’t like it, and redid it. The damage stayed within your desk.

As it gains capabilities—reading your documents, searching, calculating, preparing things to send—the worst-case scenario grows. Not because it became dangerous, but because the impact of mistakes grew along with the impact of successes. The two always grow together.

Helena experienced a small version of this: a summary with a deadline that didn’t exist. It cost her an awkward phone call. If that same summary had gone straight to the client in an automated message, without going through her, it would have cost her a client.

02 Three levels—and the middle one is where the work happens

You don’t organize this with "carefulness." You organize it with three tiers, written down beforehand. Free: what it can do without asking, because mistakes there are cheap and reversible — draft, summarize for you to read, suggest options. Requires approval: what it prepares and then pauses, waiting for your yes — anything that will go to someone else. Prohibited: what it never does, even if you ask without thinking on a busy day.

The middle boundary is what makes the difference day to day, and it’s the one almost nobody writes down. It doesn’t stop the work: it just puts you between “done” and “sent.”

Rafael wrote his in five lines. Free to do: draft captions, suggest titles, analyze campaign numbers. Requires approval: any text that will be published, any response to a customer. Prohibited: quoting a price without confirmation, promising a delivery date, using a competitor’s name.

03 Approval needs a scheduled time

"I check before sending" isn't a boundary — it's an intention. It becomes a boundary when there's a defined moment, a defined method, and a defined state: the thing is ready in a waiting place, and you say yes.

In practice, it’s what companies have done for decades with anything that goes out: there’s a test version, someone reviews it, someone signs off. AI’s time savings don’t disappear because of this—the preparation, which used to take a long time, remains automatic. The only manual part is the final click, which takes seconds.

Sônia set hers up this way: all forty notices are prepared in a list, she reads the first three and a sample from the middle, and only then sends them. Before, she wrote all forty and reread none of them because there was no time left after writing.

04 Limits don’t hold up the work—they do the opposite

It’s understandable to worry that writing down limits will make something that felt easy and quick more bureaucratic. In practice, the opposite happens: people with written limits delegate more, not less.

The reason is simple. Without written boundaries, every task becomes a new decision — “should I let it?” “what if it goes wrong?” — and in the end you do everything by hand out of caution. With the boundaries written down, you stop supervising everything within the free boundary, for good, without guilt.

Helena realized this when she counted: after writing the three bands, she started letting the AI handle internal email drafts and document triage on its own—things she used to review line by line because of a vague unease, even though they clearly fell into the cheap-to-get-wrong band.

01 The cycle has four steps, and the second is the one that’s missing

The full cycle is: do, look, correct, check again. Nothing new — it's what every professional has always done with their own work. With AI, the "doing" got so fast that the other three disappear by comparison.

When a draft takes forty minutes, spending ten minutes reviewing it seems reasonable. When a draft takes twenty seconds, spending ten minutes reviewing it seems like a waste of time—and that’s when the step disappears. Review time shouldn’t scale with production time, but it does, purely because of a psychological effect.

Rafael timed it out of curiosity: he spent an average of fifty seconds between receiving the text and publishing it. On the texts that caused problems, that number dropped to less than twenty. The error wasn’t in the AI or the text—it was in those thirty seconds that disappeared.

02 "Done" is defined beforehand, not recognized afterward

You saw this in lesson 2 as the fifth part of the request. Here, it takes on a different role: it stops being a target for the AI and becomes your review checklist. It’s the same text serving two purposes, which is why every line you write goes so far.

A good done criterion should have three to five items, all verifiable by eye in seconds. “Looks good” isn’t an item. “Every claim about amounts cites its source” is an item. The difference is that you can answer yes or no to the second one without arguing with yourself.

Helena’s case, briefly: she wrote four criteria for contract summaries — every deadline cites the clause; no legal term without an explanation; fits on one screen; anything she needs to confirm is marked. On the first try, the summary failed the first criterion. She asked for a correction to that item only, in one sentence, and the second version passed. Total: three minutes, compared with the forty she would have spent rewriting it herself.

03 AI can check — but it can’t be the only one checking

You can ask it to check the work against your list, and it’s worth doing: it catches obvious oversights, forgotten items, and formatting that doesn’t match the template. It’s quick and inexpensive.

What it can’t catch is what it invented confidently — because, to the AI, it’s on record. Helena’s contract deadline of 90 days would pass a self-review: it’s written there, it’s consistent, it fits with the rest. Only someone with the document in front of them can catch it.

Sônia uses the split that works: AI checks the form (all items included, length, tone, no empty fields), and she checks the facts (the date really appears on the official calendar, the amount matches this year’s table). It takes her two minutes, not twenty.

04 When to stop correcting

The cycle has a risk in the opposite direction: turning into an endless back-and-forth of "make it a little better". That takes more time than skipping the review and produces worse text, because each round moves the result further from what you mentally approved the first time.

The practical rule: two rounds of corrections. If it still doesn’t meet the criterion after two, the problem isn’t the result — it’s the criterion, the knowledge, or the request. Go back one layer instead of asking for a third version.

Rafael used to ask for five versions and choose the third. Now he asks for one, checks it against the items, fixes what failed, and stops. When the second one still fails, he already knows what to do: almost always, a line is missing from the dossier in lesson 3.

01 Seven steps between asking in chat and having something that runs

There’s a ladder, and it’s always the same. One: you ask and it answers. Two: it uses capabilities — reads what you attach, searches, calculates. Three: it follows a fixed recipe for a task that repeats. Four: there is a definition of done and a check that runs every time. Five: there are written boundaries and an approval step before anything goes out.

The last two are on another level. Six: it completes an entire job on its own within the boundaries, and you only step in to approve. Seven: several workstreams run in parallel, coordinated, each with its own role — what makes sense for a team today rarely makes sense for one person alone.

It’s worth noticing one thing: the first five steps don’t require any new tools. Helena climbed from one to five by writing text in notepads, using the same chat tool she already had before starting this course.

02 Where you are today—by task, never by person

The step isn’t yours; it belongs to the task. The same person is usually on step five for a task they know well and step one for another they do once a month. That’s normal and the right way to read the ladder — anyone who tries to climb "as a professional" across the board doesn’t move up at all.

The sign that a task is ready to move up is always the same: you felt the discomfort of that step. You’re tired of explaining it again (memory is missing). You’re tired of rewriting the request (a recipe is missing). You got a scare (guardrails are missing). The discomfort is the invitation, and it comes on its own.

03 One step at a time—and not every task should move up

Skipping steps is the most common way to give up. Anyone who jumps straight from one to six builds a structure that hasn’t been tested at all, gets a big scare, and concludes that “AI isn’t reliable.” It wasn’t the AI: it was a step four that never existed.

There’s another side that gets less attention: not every task should move up. A rare task doesn’t justify the cost of a standard process. A task where the conversation is the value — a negotiation, difficult advice — loses something when standardized. Deciding that a task stays on step one is a management decision, not a failure.

One short case to bring it all together. On a Friday impulse, Rafael wanted to let AI post on the chain’s social media accounts on its own. A month earlier, he had no done criterion or written limits. He didn’t do it—he stopped, built level four, then level five, and now he publishes with approval in seconds. The impulse would have cost as much as the structure took to build, plus the shock.

04 What still holds true when everything changes again

Tools will change their names, screens, and prices, and gain capabilities that don’t exist today. None of that changes what you built here, because none of these pieces is about the tool: they’re about your work.

A request in five parts will keep working. A dossier will keep working. Separating what stays saved from what’s just for the moment, the three levels of limits, and writing a done criterion in advance—all of that predates any tool and will outlast them all. Screens are what get old; the structure is what stays.

Helena sums it up this way, and it’s a good way to leave this course: she didn’t learn how to use an AI tool. For the first time in twenty years at the law firm, she wrote down how her own work gets done. AI was the reason; the procedure is what stayed.