How to Build an AI Prompt Playbook (2026)

You are reading Chapter 24 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.

An AI prompt playbook is a short document that names the three or four jobs that keep arriving at your desk and records how you handle each one: the trigger, the material, the moves in order, the check that says it is finished, and the failure you have already met. It takes an afternoon to write. It replaces the blank screen you start from every time.

This guide shows you how to find the jobs that earn a written approach, what the five lines of a play are, how a play differs from a recipe, and when to revise one rather than letting it go out of date.

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

You have eleven chapters of technique behind you. A framework, examples, decomposition, a way into long documents, a way to handle numbers without being lied to, and a taxonomy of how the assistant fails.

What you do not have is an answer to the question that actually arrives on a Tuesday afternoon. A 60-page supplier contract lands, you have two hours before the call, and nothing tells you which of those eleven chapters to reach for. So you improvise, which means the quality of your work depends on how tired you are at 3pm.

The fix is to stop organising what you know by technique and start organising it by job.

This chapter needs three things from the ones before it: the templates you wrote in Chapter 17, the quality bar you set in Chapter 22, and the failure taxonomy from Chapter 23. It also needs your calendar and your sent mail from the last 8 weeks, open, before you start.

A prompt answers a question. A template does one task with different inputs. A play is bigger than both. It is the whole approach to a job that keeps arriving, from the moment it lands to the moment you are willing to send it.

The single reusable template is the subject of Chapter 17, and Chapter 17 covers how to write and store one. A playbook is the assembled set: which jobs get a page, what goes on each page, and what order the moves run in.

Look at your recurring work and it sorts into three kinds. Something arrives that you have to understand fast, like a contract, a board pack or an analyst report. Something is due that you produce to a standard, like the monthly update or the deal review. And something has to be decided and then defended to a person who will push back. Most professionals are heavier on one of the three than they think, and which one dominates your week tells you which of the last eleven chapters you will lean on hardest.

Not every recurring job earns a page. The test has three parts: it happened at least 3 times last quarter, a bad version costs you real time or credibility, and you rebuild the approach from scratch each time. All three, or it stays off the list.

The third part is the one that keeps the list short. Plenty of things you do weekly you already do fluently, in a groove, without rebuilding anything. Those need nothing from you. The jobs worth writing down are the ones where you start from a blank screen every time even though you have done it eleven times before.

Three or four plays, not twelve. A playbook with twelve plays is a document nobody opens, and the first person to stop opening it is you.

Do not do this from memory. Memory hands you the unusual month and hides the ordinary one. Open your sent mail and your calendar for the last 8 weeks and read them as a stranger would, looking only for the thing that appears again and again. Most people are surprised by at least one entry on that list, and the surprise is usually the most useful play in the book.

A play is not a longer prompt. It is the small amount of thinking you do every time this job arrives, written down once, so that you never have to do it while tired. One page, five lines, and none of those lines is a prompt. The prompts stay in the template library you built in Chapter 17. The play tells you which ones to take down, and in what order.

Write the trigger as a moment, not a topic. "A supplier proposal lands in my inbox and I have to recommend yes or no" is a trigger. "Supplier analysis" is a filing label, and a filing label will not occur to you at the moment you need it.

Duller and more useful than it sounds. Almost every bad run of a job begins with starting it without the thing you needed: last year's version, the actual numbers rather than the summary of them, the policy you are meant to be checking against. Write the list once.

Two or three moves, usually. Each one can point at a template you already wrote, so the play stays short rather than becoming a procedure manual. If a play needs eight moves, it is two plays.

Take the quality bar you set in Chapter 22 and write it as a question you can answer yes or no in under a minute. Could a reader answer the 3 questions my CFO always asks, from this page alone. This is the line almost everybody leaves off, and without it you are finished when you run out of energy, which is not the same thing as finished.

Chapter 23 gave you the general taxonomy. A play carries only the one that has actually bitten you on this job. On contracts it is usually invented specificity, a clause number that reads perfectly and is not in the document. On anything with a table in it, it is arithmetic done inside prose. One line, naming the failure and naming the 30 seconds that catches it.

The difference between the two is not length. It is whether the thing still works on an input you were not thinking about when you wrote it.

A recipe is one task, one prompt, one good result, kept because it worked on Thursday. It takes a minute to write, which is why people write them. It quietly assumes the next input looks like the last one, it says nothing about what to do when the answer comes back wrong, and it breaks the first time the input arrives in a different shape.

A play was written from 2 or 3 instances that genuinely differed. It names the kind of material rather than one specific document, it says what going wrong looks like, and someone else can run it and land on your standard instead of their own. It costs you a second and third run before you can trust it, and it is tempting to over-write, at which point nobody reads it again.

The test: hand it to a competent colleague along with a file you have never seen, and see whether they come back with your standard or theirs.

Here is a whole play, written out. The job is a supplier proposal that arrives when you have to recommend yes or no to the executive team inside two days. It happens 5 or 6 times a year, it always matters, and you have always started it from nothing.

The trigger is the proposal landing. The material is the proposal itself, the incumbent contract, and the 3 numbers your finance director will ask about, all in front of you before anything starts. That second line alone removes the most common bad version of this job, the one where you get 40 minutes in and find you were reading an out-of-date pricing annex.

Then three moves. Pull out what the supplier is committing to, what is conditional and what is simply absent, working from the document rather than from the assistant's general knowledge of supplier contracts. Check every figure and every date separately, out of the prose. Then make it argue against your own emerging recommendation and see which objection you cannot answer, which is the move most people skip, because by then they have decided.

The check: could someone read my one page and answer the 3 questions the executive team always asks. Yes or no, in under a minute. The known failure, one line: it invents clause numbers that read perfectly, so every clause reference gets matched against the document before this leaves my desk.

Read that back and notice what happened. Chapters 18, 19, 21, 22 and 23 stopped being five techniques you might remember to use and became five lines you will perform in order, because they are written on the page in front of you.

Notice too that one move in a real version of this play is not an AI move at all. At some point you telephone the account manager. Write that step in. A play that pretends every step belongs to the assistant is the fastest way to use the assistant on the one part it is worst at.

You cannot feel whether any of this is working, and that is a measured finding rather than a character flaw. In 2025 a research group ran a trial with 16 experienced open-source developers across 246 real tasks on codebases they had worked on for years. Before starting, the developers expected AI tools to make them about 24% faster. Measured, they were 19% slower. Afterwards they still believed they had been roughly 20% faster.

Those are software developers on their own repositories, and none of it transfers directly to your inbox. What transfers is the size of the gap between how it felt and what happened, in people who were experts at the underlying work.

Which is why a play carries a date and one line saying what it produced the last time you ran it. A play with no date is a claim with no evidence, and in 6 months you will not remember whether it predates the tool you now use. The line of proof is the only defence against keeping a play out of affection.

This is deliberately not a weekly ritual. A review you schedule is a review you cancel. Four triggers are the whole maintenance regime, and you read down until one is true.

A play failed the last time you ran it, so fix the play rather than just the prompt. The job itself changed, so rewrite the check first and then the moves. You have not opened it in 8 weeks, so it was the wrong job and you delete it. It runs identically every time and you no longer read it, which means it is ready to hand over.

Prune harder than you add. Four plays you actually run beat twelve you have, and a playbook grows by deletion more than by writing.

The encouraging half is more specific than the usual claims. A field experiment with 776 professionals at Procter & Gamble, published in March 2025, found that people working outside their usual area did poorly on their own without AI, but with AI they reached the level of teams that included someone experienced in the work. The authors' reading is the part that matters here: the tool let less experienced people reach a standard that used to need a colleague or a supervisor. Writing down how you do a job well, so it runs the same way on a bad Thursday, is you doing that for yourself.

And the counterweight, because this course does not sell. Economists tracking the Danish labour market found that 2 years after ChatGPT launched, with adoption widespread and workers reporting real time savings, there were precise nulls on earnings and recorded hours, with any effect ruled out above 2%. The work visibly reorganised. The measured payoff did not show up where everyone expected it. Adoption is not benefit, and being able to name exactly what you now do differently, then check whether it held, is the version of this that holds up when your boss asks a serious question.

Before: you know a great deal about prompting and you still start every recurring job from a blank screen, so the result depends on what time of day the work lands.

After: you have a written playbook. Your recurring jobs are named, three or four of them have an approach recorded, each one has a check that tells you when it is finished, and a rule keeps the whole thing from going stale.

That closes AI 202. The track set out to give you the framework behind every good instruction and the ways to catch AI when it is wrong, and this chapter is where those eleven chapters stop being techniques you might remember and become lines on a page you perform in order. If you want the techniques in one reference, the 6-step AI prompt formula covers the core of them, and these 14 essential ChatGPT prompt templates give you a starting set for the moves in each play.

Two things carry into Chapter 25: your playbook, with three or four plays written and dated, and the one play you marked as the job you would most like to stop sitting through. Chapter 25 opens AI 303, which is about connecting AI to your tools and building automations that run while you are not looking. It works on a job whose trigger, steps and check already exist in writing, and that marked play is it. More guides like this one are in the AI Central Library.

A short document that lists your recurring jobs and records a written approach for the three or four that matter most. Each job gets one page holding five lines: the trigger, what has to be in front of you, the moves in order, the check that says it is finished, and the failure you have already met on that job.

Three or four. A playbook with twelve plays is a document nobody opens, and the first person to stop opening it is you. Prune harder than you add: four plays you actually run beat twelve you have.

A template does one task with different inputs. A play covers a class of work: a trigger, the material, several moves, a check and a known failure. A template is written from one good result and breaks when the input arrives in a different shape. A play is written from 2 or 3 instances that genuinely differed, and the moves in it can each point at a template you already wrote.

Three tests, and it has to pass all three. It happened at least 3 times last quarter, a bad version costs you real time or credibility, and you rebuild the approach from scratch each time. Find the candidates in your sent mail and your calendar from the last 8 weeks rather than from memory, because memory over-reports the unusual month.

When one of four things is true. It failed the last time you ran it, so fix the play rather than the prompt. The job changed, so rewrite the check first. You have not opened it in 8 weeks, so delete it. Or it runs identically every time and you no longer read it, which means it is ready to hand over to an automation.