How to Break a Big Task Into AI Prompts (2026)

You are reading Chapter 16 of 36 in the AI Bootcamp, AI Central's free course for professionals who want to use AI in their own work. The chapters build on each other, so if this is your first one, Chapter 1 is the place to start. All of them sit on the AI Bootcamp page, and three more arrive every week.

Breaking a big task into AI prompts means running it as a sequence of short messages instead of one long request, where each message does one job and hands its answer to the next. The reason to do it is not that the tool cannot hold the whole job. It is that a single reply gives you nowhere to correct a decision before it is set in 900 words of prose.

This guide shows you why one enormous request comes back thin, runs the 5-step sequence on a real board update, and gives you the test for where to cut a job and where not to.

We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.

Here is the situation this chapter exists to fix. You have a quarterly board update to write, you have finally cleared 40 minutes, and you put the whole thing into one message: the quarter's numbers, the miss on renewals, where you stand against two competitors, and the headcount you intend to ask for. Back comes 900 words that are fluent, correctly ordered, and say nothing you could argue with.

Nothing in it is wrong. That is the frustrating part. You cannot fix it with an edit, because there is nothing specific enough to edit, and you cannot say what you wanted instead, because what you wanted was for it to have decided something. If you cannot find a sentence in the output to disagree with, the ask was too big.

Look at what you actually asked for. Six separate jobs in one sentence: recall the facts, work out what the story is, decide the order, write the difficult section, anticipate the pushback, and produce polished prose. All six at once, with no point at which anything could be checked with you.

The fix is not a better paragraph of instructions. It is the same work run as a short sequence, where each message does one of those jobs and hands its result to the next. You read what comes back at every stage, correct it in a line or two, and that correction travels forward instead of being discovered in the meeting.

It costs you about 5 messages instead of 1, and about 20 minutes instead of 3. That is worth paying on work with your name on it that a room of people will read closely. It is not worth paying on a 200-word email.

This chapter assumes you already have a prompt framework and the habit of asking the assistant to reason before it answers. If you do not, Chapter 15 covers asking it to reason before it answers, and it is the chapter directly behind this one.

This is not laziness and it is not a badly worded prompt. Three separate things happen when a request gets large, and knowing which one is at work tells you where to cut.

Ask for 6 things and the effort divides roughly evenly between them, because nothing in your request said which one carries the argument. The throwaway point and the load-bearing point get the same three sentences. A human writer solves this by deciding what the piece is about before writing a word. Unless you make that decision a step of its own, it never gets made.

Every one of these assistants will read far more than it will write in one reply. The same tool that reads a 200-page contract without difficulty produces a few thousand words in one go, and long-form benchmarks through 2026 keep finding the same pattern: models fall short of explicit length targets, and the shortfall grows the longer the target gets. Ask for the whole document and you are asking it to compress. Ask for one section and you are asking it to write.

A long input is not read evenly across its length. Chroma's context rot study tested 18 frontier models and found every one of them getting less accurate as the input grew, with the steepest fall in the range people reach when they paste in everything they have. In a separate evaluation of 4 leading models, accuracy on a simple two-step question dropped from between 82% and 99% to between 22% and 70% once that same question sat inside 32,000 tokens of surrounding material.

One part of this advice has genuinely weakened, and it is worth saying so. Models reason internally now in a way they did not three years ago, and much of the step-by-step scaffolding people used to type by hand is done for you. That is real progress, and it is Chapter 15's territory. What has not changed is the rest. Context windows in the millions did not stop the degradation, and a larger window has never cured the thinness.

Anthropic's own current guidance says it plainly: the value of splitting a task is no longer mainly about fitting inside a context window, it is about letting the model do its best work inside one narrow set of requirements. You are not breaking work up because the machine cannot hold it. You are breaking it up so you can see and correct each decision before it is set in 900 words of prose.

Here is the same deliverable run properly, start to finish. The sequence itself is the thing worth keeping, because it transfers to a competitive review, a customer research synthesis, or any strategy memo you have to defend.

Give it what you have: the numbers, last quarter's deck, the renewals report. Ask for an inventory rather than a draft, meaning which facts are in this material, which of them a board would consider load-bearing, and what is missing that a board would expect to see. What comes back is a list. Reading it takes 2 minutes, and it usually surfaces at least one gap you would otherwise have met for the first time in the meeting.

Working only from that inventory, ask for the section list and, for each section, the single sentence it has to land. This is the step people skip and it does more work than any other, because it is where the document acquires an argument rather than an outline. Cut sections, reorder them, rewrite two of the sentences in your own words. What you approve here is what every later step inherits.

Draft one section in full, and make it the hardest one, which in this case is the renewals miss, at the depth you would defend in the room. You come out with two things: a section you can actually use, and a specimen of what good looks like in this document, which the remaining sections can be told to match. It is also the section most likely to change the structure, which is why it goes before the easy ones.

Give the drafted section back and ask what the most sceptical person at the table goes after first, and what they would ask that this section does not answer. You want objections, not a rewrite, so say so. Chapter 21 is entirely about getting genuine criticism out of something whose instinct is to agree with you. Here you need one honest pass, not the full treatment.

Revise against those objections, and only then write the remaining sections against the approved skeleton and the approved specimen. Notice what the order protected. By the time any easy section gets written, the structure is settled, the difficult section exists, and the objections are already known.

Two things make this a chain rather than five separate messages. You read and correct in the gaps, because firing all 5 without looking means you have only typed more. And you run the whole sequence in one conversation, so each step can see the last.

The sequence behind it is 5 words long: gather, structure, draft, critique, revise. Write those 5 down.

Splitting a job is easy. Splitting it in the right places is the actual skill, and a bad cut is worse than no cut, because you end up assembling pieces that were each written as though the others did not exist.

The most common bad cut is by section. You ask for section 1 in one message and section 2 in the next, and you get two documents: two openings, the same background explained twice, two different registers, and no argument running through either. Nothing in the second message knew what the first one decided.

The cut that works falls where a decision gets made. Gather and structure are separate steps for exactly that reason: what the inventory turns up changes the skeleton, and the skeleton changes everything after it. The practical version of the test is to ask whether you would give this piece to a capable colleague and expect something finished back. If the answer is yes, it is a step.

Two things veto a cut, and the veto wins. If the halves have to stay consistent with each other, in register, in a running argument, or in a comparison across several items, keep them in one step so both are in front of it at the same time. And if you could not judge a step's output in about 2 minutes, it is not a checkpoint, it is a delay: you have bought waiting time and no control.

You can also cut too far. Fewer than 3 steps and you have not really split it. Past 8 or so, a chain starts losing the thread it began with. Each step carries a little less of the original intent, a mistake in step 3 travels into everything downstream, and by step 10 you are administering a process instead of writing a document. Five is a comfortable size for a real deliverable. If you find yourself at 12, you have cut along seams that were not there.

None of this applies to a 200-word email. One well-formed prompt handles that, and reaching for a 5-step sequence there is a way of looking busy. Keep it for work that has a shape: sections, an argument, and someone who will read it closely. If your single prompts are not landing in the first place, fix that first with the 26 principles of prompt engineering.

The chapter closes with 3 exercises that take about 25 minutes: name the big job, make the first cut, then run steps 1 and 2 only.

Before: you clear 40 minutes, put the whole deliverable into one message, and get back 900 competent words you cannot edit into something better, because there is nothing specific enough in them to edit.

After: you can tell which jobs are too big to ask for in a single message, you run one of them as a sequence of 5 steps where each step takes the last one's output as its input, and you place the cuts so the pieces assemble into one document instead of a pile.

Two things carry into Chapter 17: the sequence you actually ran, and the step names you wrote down in your own words. Chapter 17 turns those names into prompts you can run again, so a chain you worked out once takes 5 minutes next quarter instead of an afternoon. If you want a one-page reference to keep beside you while you work, start with our free AI cheat sheets, and the rest of the syllabus lives in the AI Central Library.

Three things happen at once. Effort spreads roughly evenly across everything you named, so the load-bearing point gets the same three sentences as the throwaway one. A single reply has a length ceiling far below what the tool can read. And a long input is not read evenly across its length, so accuracy falls as the material grows.

Around 5 for a real deliverable. Fewer than 3 and you have not really split it. Past 8 or so a chain starts losing the thread it began with, a mistake in an early step travels into everything downstream, and you end up administering a process instead of writing a document.

Cut where a decision gets made, meaning anywhere the answer could change what you ask for next. The practical test is whether you would hand that piece to a capable colleague and expect something finished back. Two things veto a cut: halves that must stay consistent with each other, and any step whose output you could not judge in about 2 minutes.

No. Context windows in the millions did not stop accuracy falling as the input grows, and Anthropic's current guidance puts the value of splitting a task in letting the model work inside one narrow set of requirements rather than in fitting the material. Models do reason internally now, which removes some of the scaffolding people used to type by hand, but it does not remove the rest.

Gather, structure, draft, critique, revise. Message 1 asks for an inventory of what you have and what is missing, message 2 agrees the section list and the sentence each section has to land, message 3 drafts the hardest section in full, message 4 asks what a sceptic attacks, and message 5 revises against those objections before the easy sections get written.