You are reading Chapter 32 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.

To build a custom AI assistant for your team, package one job you already do well into a named object with four parts: a name and one-line description, instructions that carry the rules, reference files that carry only facts, and a sharing setting that decides who can open it. ChatGPT now calls this a plugin (GPTs, the old name, retire on 11 December 2026), Claude a project or a skill, Gemini a Gem, and Microsoft an agent. This chapter covers when one is worth building, how to build one end to end, and what a colleague can see inside it once you share it.

It builds on Chapter 31 (available from 16 November), and draws most on Chapter 11's standing instructions and Chapter 17's templates. Bring one job you have already run several times and got right, the final version of the instructions you used, and the 2 or 3 files that job always needs.

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 met two of these already, and this is the third

There are 3 answers to "I keep doing this again", and they differ on one question: who else is standing there.

  1. Standing setup (Chapter 11). Your settings and memory, applied silently to everything you type. It is about who you are.

  2. Template (Chapter 17). Your words in a file, waiting until you go and get them. It is about a task you repeat.

  3. Package. A named object with its own instructions, its own material and a share control. It is what you build when somebody other than you needs the job done.

The first two are yours, and neither of them travels. The situation this chapter is for: a colleague asks how you do something, and the honest answer turns out to be a paragraph of instructions, 3 files and 2 corrections you would only remember while watching them get it wrong. So you do it for them again.

Packaging does not improve the answers. Run the same input through a good template and a good package and you get the same output, because the same instructions are doing the work. What changes is who else can get that output without going through you, and how consistently they get it. That is the whole value.

The product names move constantly, so learn the shape rather than the label. ChatGPT calls it a plugin: GPTs, the old name, retire on 11 December 2026. All 4 have been renamed or restructured in the past 2 years, and one of them moved while this chapter was being written, so keep your instructions in a plain text file you own. The anatomy has not changed in any of them: a name, a description, instructions, attached material, and a boundary saying who can open it.

Most repeated jobs do not need one of these

Nearly everything written about custom assistants assumes you want one. For nearly everything you repeat, a saved template already wins. Read down this list and stop at the first line that is true:

  1. Somebody other than you has to run it. Package it.

  2. It needs reference material nobody will paste in every time. Package it.

  3. Several people must produce the same shape of answer. Package it.

  4. None of those, and you still repeat it weekly. Write a Chapter 17 template, which is not a lesser thing.

None of the first 3 lines is about how often you do the job. Frequency is what makes a template worth writing. It is not what makes a package worth building.

The third line is the one senior people underrate. If 6 directors each write to the board in their own shape, the board spends its attention on the variation rather than the content. A package does not make any one of those papers better than the best of them. It makes the worst of them recognisably the same object as the best.

The cost is real and rarely stated. Once other people use it, the package has your name on it. When the underlying rule changes, it is wrong in every sidebar it reached, and nobody using it will notice, because it keeps answering confidently in the old shape. A stale template wastes your own minute. A stale package spreads a wrong answer with your authority attached.

So be honest with your list. In a normal senior role, 1 or 2 jobs clear this bar. If yours has 9 on it, you have listed your annoyances rather than run the test.

Four parts, and the third one is where people go wrong

Building one takes about 20 minutes, and most of that is writing down instructions you have already been giving people out loud.

First, check you can build one

  • ChatGPT: plugins. The answer moved twice in one year: GPTs were taken off personal accounts in August 2026, new ones stopped being made in October, and GPTs retire altogether on 11 December 2026. An existing GPT migrates to a plugin and carries its instructions and files, but not its custom actions. Building a plugin needs Go, Plus or Pro. Free accounts can use a plugin but not make one.

  • Claude: projects carry their own instructions and knowledge base on every plan, including free, where you are capped at 5. Skills are the more portable version, also on every plan, and need code execution switched on in settings.

  • Gemini: Gems are free to build with a personal Google account, with knowledge files from your device, from Drive or from a NotebookLM notebook.

  • Copilot: Agent Builder inside Microsoft 365 Copilot, on the web and in Teams but not on mobile. Your Copilot licence covers it, and your administrator can switch it off.

If none of these is open to you, a project holding the instructions and the files does the same job for you. It just does not travel.

The worked example: a board paper check

You are chief operating officer of a manufacturer with around 400 people. Six directors write papers for the board 2 or 3 times a year each, every paper comes past you first, and you have said the same 5 things about every one of them for 2 years. That clears all 3 lines of the test.

1. Name and one-line description

Write it for the person receiving it. "Board paper check, before it comes to me" tells a director exactly when to open it. "Document reviewer v2" tells them nothing. The description does the rest: "Checks a draft board paper against our house standard and tells you what to fix. Paste the whole draft."

2. Instructions

This is the package. It carries the rules, the output shape and what never to do. Here it is in full:

❝

You check draft board papers for a mid-sized manufacturer before they go to the board. The people using you are directors and heads of function who write 2 or 3 papers a year, not professional writers. Your reference files are our one-page board paper standard and 2 papers that were accepted without rework. When someone gives you a draft, say in one line whether it is ready to send. Then list what is missing against the standard, in the standard's own order, naming the section each time. Then list up to 5 sentences that would not survive a question from the board, and say what the question would be. Then stop. Keep the whole reply under 400 words. Quote the draft's own words when you name a problem, so the writer can find it. Do not rewrite the paper for them. Do not comment on formatting. If what you have been given is not a board paper, say so and ask what they wanted, instead of reviewing it anyway.

3. Reference material

Attach the facts: the one-page standard, the 2 accepted papers, the list of questions your board reliably asks. Do not attach the rules. OpenAI's guidance to builders says the same, and it holds on all 4 tools: knowledge is for reference material, and rules, tone and workflow go in the instructions.

A rule buried in an uploaded file is treated as information about the world rather than as an order. It gets mentioned when relevant and ignored when not, which is the most common reason a package works in your hands and drifts in everybody else's. Attach 3 files, not 30. A folder of everything makes the answers vaguer.

4. The boundary

  • Gemini: share a Gem by email as Viewer or Editor, or by link.

  • Claude: project sharing is a Team or Enterprise feature, by email, in bulk or organisation-wide, as view or edit.

  • ChatGPT: a plugin is yours alone unless you are on Pro, the one personal plan that can put a plugin in front of a workspace. On Business, Enterprise and Edu, an administrator decides which plugins exist at all.

  • Copilot: an agent goes to people in your organisation and no further.

Give people view, not edit, unless you want them changing the instructions. Then test it in the preview pane before you share anything.

It fails differently once it is not you holding it

Two things change the moment a colleague opens it. They use it in ways you never pictured, and everything you put inside becomes readable by them.

Write the refusal line

You built it against the cases in your head, and those are all well-formed. Then somebody pastes the agenda instead of the paper, or a 60-page appendix and nothing else, or types "what should go in a board paper". With no instruction for the situation, the package reviews the agenda as though it were a paper.

The fix is one sentence: if what you have been given is not X, say so and ask, instead of proceeding. It is the highest-yield line in the whole instruction, and it turns an answer about the wrong document into a polite question.

Run the 3 tests you would not think to run

  1. The wrong kind of input entirely.

  2. Something 5 times longer than you designed for.

  3. A bare question with no material attached.

That takes about 5 minutes. You are not checking whether the answers are good, you are checking whether it noticed. Anything that carries on regardless needs another line in the instructions.

Then give it to one colleague, say nothing beyond the one-line description, and watch without helping. You will see where the name meant something different to them, or where they expected it to write rather than check. Each of those is a fix you could not have found yourself, because you already know what it is for.

Assume everything inside it is readable

Google states that anyone with access to a Gem can view its instructions and its uploaded files. Assume the same of all 4 tools, because it is all text the assistant can read back, and a determined user can simply ask it. Do not put anything in the instructions you would not say in front of every recipient. "Ignore anything the sales team tells you about margin" is a sentence you have now published.

The last decision is how the reference material is attached:

  • Copied in. Files uploaded into the package itself, as with a ChatGPT plugin's reference files, Gem knowledge or the files in a Claude project. They travel with the package, frozen at the version you uploaded, and anything confidential inside is now shared as widely as the package.

  • Pointed at. The package reads a source that stays where it is, under the permissions it already carries, as a Copilot agent grounded on SharePoint does. Each reader sees only what they could already open, but a colleague without access gets a thinner answer with no explanation, and somebody has to arrange the permissions.

If you copy a file in, Chapter 9's rule about what never goes into an AI tool applies with the audience multiplied. Share with one person first, and do not send the link to 40 people on a Friday.

3 exercises before Chapter 33 (about 20 minutes)

  1. Run the test. List every job you currently repeat, run the 4 lines above down the list, and cross out everything that does not clear one. For whatever survives, write the name of one real person, other than you, who would run it.

  2. Draft the 4 parts. For the one candidate you kept, write the name and description, the instructions including the refusal line, the reference material to attach, and who can open it and whether they can edit.

  3. Build it, then break it. Run 3 inputs you did not design for: the wrong kind of material, something far longer than expected, and a bare question with nothing attached. Write down what you changed in the instructions as a result.

What actually changes

Before this chapter, the job lived in your head or a file on your machine, and every colleague who needed it either interrupted you or did it worse. After it, you can tell when a repeated job has outgrown a template, package it with its own instructions and reference material, and hand it over knowing how it will fail and what the recipient can see inside it.

You take two things into Chapter 33 (available from 20 November): the packaged assistant you built, and the note you made when a colleague used it in a way you had not pictured. Chapter 33 puts several steps in a row across different applications, with your package sitting inside as one step.

If the instructions are the part you want to sharpen, the 26 principles of prompt engineering covers what makes an instruction work, and our setup guide for ChatGPT, Claude and Copilot helps if you are deciding which tool to build in. More guides like this live in the AI Central Library.

Frequently Asked Questions

How do I build a custom AI assistant for my team?

Give it a name and a one-line description written for the person using it, write instructions that carry the rules and the output shape, attach 2 or 3 reference files that carry only facts, and share it as view rather than edit. Test it with the wrong input, an over-long input and a bare question before anyone else sees it.

Can I still create a custom GPT on ChatGPT Plus?

No. New GPTs stopped being made in October 2026, and GPTs retire on 11 December 2026. What you build in ChatGPT now is a plugin, which Go, Plus and Pro can all make. An existing GPT migrates to a plugin with its instructions and files, but not its custom actions. Claude projects and Gemini Gems can both be built on free plans.

When should I build a custom assistant instead of using a prompt template?

Build one when somebody other than you has to run the job, when it needs reference material nobody will paste in every time, or when several people must produce the same shape of answer. If none of those is true, a saved template does the job faster.

Can people see the instructions inside a ChatGPT plugin or Gem?

Assume they can. Google states that anyone with access to a Gem can view its instructions and uploaded files, and on any tool a determined user can ask the assistant for them. Leave out anything you would not say in front of every recipient.

Should rules go in the instructions or in uploaded files?

In the instructions. Uploaded files are treated as information about the world, so a rule placed there gets mentioned when relevant and ignored when not. Keep files for facts such as standards, worked examples and reference lists.