You are reading Chapter 17 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.
A reusable AI prompt template is a prompt that already worked, with the parts that were only true that day replaced by named slots anyone can refill. Building one takes about a minute, because you are not writing from scratch. You are annotating a prompt you have already sent.
This guide shows you how to tell a template apart from a standing instruction, how to split a working prompt into its fixed and swappable halves, the four lines worth writing down in the half-minute after it lands, and where to keep it so you find it again.
We publish AI systems like this for 300,000+ senior professionals at AI Central, and more of them live in the AI Central Library.
Why your best prompts keep disappearing
You have written good prompts. Some of them took three or four rounds of fixing before the answer came back right. Almost all of them are now sitting in a conversation you are never going to scroll back through.
So next month you write the same prompt again, slightly worse, from memory.
The waste is not the typing. It is the thinking. A prompt that took four attempts contains four corrections you will not remember making, and rebuilding it from scratch means paying for those corrections a second time.
This chapter needs one thing from you before it works: a prompt from the last four chapters that produced something you actually used. If you have not been through them yet, Chapter 16 is the one immediately before this. Go and find the prompt before you read on. If you cannot find it, that is the chapter making its own point.
A prompt that worked is an asset, and you keep throwing it away
There are two ways to stop repeating yourself, and they are not interchangeable.
A standing instruction is set once and then applies by itself to every conversation. It works whether or not you remember it exists, which is the whole value. Chapter 11 covers standing instructions, memory and project-scoped context, so this chapter does not cover them again.
A template is the opposite kind of object. It sits in a file doing nothing until you deliberately go and get it. That sounds like a weakness and it is the point, because something that only fires when you ask for it can be as narrow and opinionated as you like.
You would never set "always give me six bullets, then three questions to ask before the next call" as a standing instruction. It would ruin every casual question you asked for the rest of the year. As a template you reach for on vendor proposals, it is exactly right.
So the test is one line. Would you want this behaviour on for every conversation you have? If yes, it is a standing instruction and you should go back and set it. If no, and it is still something you do more than once, it is a template.
Go and find one prompt from the last four chapters that produced something you actually used. That is what the rest of this chapter works on.
Every prompt that worked has a part that changes and a part that never does
The skill is not writing a template from nothing. It is looking at a prompt you already wrote and seeing which half of it was only about today.
Take a real one. Say you asked for this, and it gave you something you took into a meeting:
"You are advising a chief operating officer. Below is a proposal from a logistics vendor called Harlow. We are deciding whether to move European returns handling to them in Q1. Pull out what they are actually committing to, what is conditional, and what is missing. Give me six bullets, then three questions I should ask before the next call. Keep it under 200 words and do not soften anything."
The vendor changes. The document changes. The decision changes. Everything after that does not: the three things to pull out, the six bullets, the three questions, the word limit, the instruction not to soften. That second half is the part that took three attempts to get right, and it transfers untouched to every proposal you will ever read.
Mark the four things that moved and leave the rest alone, and you get this:
"You are advising a [role]. Below is a [document type] from [source]. We are deciding whether to [decision]. Pull out what they are actually committing to, what is conditional, and what is missing. Give me six bullets, then three questions I should ask before the next call. Keep it under 200 words and do not soften anything."
That is the whole conversion, and notice how little of it is writing. Here it is as four passes.
1. Copy the prompt that worked
Paste it into a blank note. Take the version that produced the answer you used, not the version you have in your head.
2. Underline what was only true today
Read it line by line and ask of every sentence whether it would be word for word identical next time. Anything that would not be is a slot.
3. Replace each with a bracketed name
Square brackets are worth using because they are readable and no assistant treats them as special syntax. Name the slot, not the value: [vendor name] tells the next person what goes there, and [Harlow] tells them nothing. A leftover example is the most common way a template gets sent out with somebody else's vendor name still in it.
4. Refill it for a different case
This is the pass people skip, and it is the one that catches the detail you left hard-coded by accident. Refill the template for a genuinely different case and send it. If the answer comes back wrong, something specific to the original job is still sitting in the fixed half, and you will find it in about ten seconds by comparing the two answers. A template you have used only once is a draft.
Two habits make the result hold up. Add a short line at the top listing the slots to fill, so you cannot run it half-filled by accident. And if your prompt carried a worked example, keep the example inside the template. It is fixed, not swappable, and stripping it out to save space is how a template quietly stops producing what the original did.
The moment to save a prompt is the moment it works
Not at the end of the day, and not when things calm down. The window is roughly 30 seconds wide, and it closes while you are still reading the answer.
Everything in this chapter fails at one point, and it is not the writing. It is that nobody captures a prompt while it is still on screen. You read a good answer, you think that is exactly it, and then you use the answer, because using the answer is the thing you actually came to do. By the evening the prompt is under nine follow-ups and you could not pick the winning version out of the near-misses if you tried.
So make the trigger the reaction, not the calendar. The thought "that is exactly it" is the signal to stop for half a minute and copy the prompt out, before you touch the output. It feels like an interruption exactly once.
Copy the last version, not the first one you typed. The prompt that worked is your opening message plus every correction you made in the follow-ups, folded back in. If the fix happened across three messages, write it as one prompt now while you still remember what the three fixes were.
Then add three short lines around it. A name, in the words you would actually search for. One line saying when to reach for it, which is what makes it findable by somebody who is not you and by you in four months. And one line on what it produced the day it worked, which is how you decide later whether it is worth keeping.
One rule keeps the file from filling with prompts you never open. A prompt earns its place the second time it works, not the first. Copy it out the first time either way, but keep the second use as the bar for promoting it into the library proper, because a file full of prompts you used once looks exactly like a file full of prompts you will never use again.
Keep them where you will actually look
All four assistants now offer somewhere to put a prompt. Most of what they offer is standing context rather than a template you invoke.
Inside the assistant
A project, a Gem, a saved prompt: storage built into the tool you are already typing in. There is nothing to switch to at the moment of need, and files and knowledge can travel with the prompt. It is locked to one assistant, awkward to share with colleagues, and easy to forget exists at all, and a feature you forget about saves you nothing.
ChatGPT has custom instructions on every plan, Projects on free and paid, and custom prompts that its docs say appear in the composer as /prompts commands. Claude has Projects, where instructions written into a project apply to every conversation inside it. Gemini has Gems: a name, instructions and optional knowledge files, free with a personal Google account, and Google says to write the instructions as persona, task, context, format. Copilot has saved prompts in Prompt Lab inside Copilot Chat, tied to your Microsoft 365 licence.
Projects and Gems are containers you work inside, and instructions written into them apply automatically to everything you do there. That is Chapter 11's idea with a boundary drawn around it, and it is genuinely useful. It is not a template you reach for. Of the four, only Copilot's saved prompts and ChatGPT's custom prompts are a place you store a prompt and invoke it later.
A document you own
One file in the notes app or doc tool you already have open, one prompt per heading. It is searchable with the search you already use daily, it works identically no matter which assistant you switch to, and it costs one paste to use and one paste to add.
It survives you changing tools, changing jobs, or losing a subscription tier. And you can send a colleague the whole thing in one link, which no native feature on this list does gracefully. The cost is two extra clicks every time you use one, and nobody maintains it for you, so it decays if you never prune.
Use the native feature where it is genuinely better, and it sometimes is. If you do the same job every week inside one assistant, a project or a Gem with the instructions written in saves you the paste. Just be honest about the failure mode: a feature you have to remember exists is a feature you will stop using in three weeks, and the prompt you saved there is then more lost than if you had never saved it.
Naming is what makes the document work, and the rule is one line long. Name the prompt after the moment you need it, not after the technique it uses. "Vendor proposal to decision brief" gets found. "Summariser v3" does not, because in month 3 you will search for the word proposal and never for the word summariser.
Keep the whole thing to one file for as long as you possibly can. Folders add a decision every time you save something, and a decision at the moment of capture is what stops you saving. When one file genuinely stops working, split it once, by the kind of work rather than by the tool.
The chapter closes with 3 exercises and about 15 minutes of work: find the prompt, split it into the half that was only true that day and the half that is identical every time, then write the finished template, name it in the words you would search for, and write down the one place it now lives.
What actually changes
Before: your best prompts sit in chat histories you search by keyword and mostly by luck, so you rebuild them from memory every few weeks, slightly worse each time.
After: you can look at a prompt you wrote once and see which half of it repeats, mark the slots so anyone can refill them, copy it out in the half-minute after it lands, and find it again months later.
Two things carry into Chapter 18: three templates with their slots marked, and the one place you keep them. Chapter 18 hands you the job most worth templating next, reading a long document, and you will want somewhere to put the prompt when it works.
For more on what belongs in the fixed half, the 26 principles of prompt engineering covers the instruction patterns worth keeping, and the ChatGPT-5 prompt optimizer is for tightening a prompt before you template it. If you want a one-page reference beside you while you work, keep our AI Dictionary handy, and the rest of the syllabus lives in the AI Central Library.
Frequently Asked Questions
What is a reusable AI prompt template?
A prompt that already worked, with the parts that were only true for that one job replaced by named slots in square brackets. The instructions stay fixed and you refill the slots each time you use it. Square brackets work because they are readable and no assistant treats them as special syntax.
What is the difference between a prompt template and a standing instruction?
A standing instruction is set once and applies by itself to every conversation. A template does nothing until you go and get it, which is why it can be far more specific. The test is whether you would want that behaviour running on every conversation you have. If not, and you still do it more than once, it is a template.
When should I save a prompt that worked?
The moment it works, before you use the output. The window is roughly 30 seconds wide and it closes while you are still reading the answer. Save the last version, the one after your corrections, rather than the first message you typed.
What should I write down with the prompt?
Four lines. A name in the words you would search for, the final prompt with its slots already marked, one line naming the moment you should reach for it, and one line on what it produced the day it worked.
Where should I keep my AI prompt templates?
One file in the notes app or doc tool you already open every day, one prompt per heading. Projects, Gems and saved prompts inside an assistant work too, but they are locked to that assistant and easy to forget. A document you own survives changing tools, and you can send the whole thing to a colleague as one link.

