Pack the brief into a Gem, a GPT or a project. A third repeat is the signal that the task is regular. Your own agent keeps the whole brief, so nothing gets lost.

What would you tell him?

Nothing personal in a public agent: files and instructions can leak. A 2025 study: over 98.8% of public GPTs gave up their instructions to tricky prompts. A passport doesn’t belong there.

Why is it better to keep only the tools you need?

Fewer ways to go wrong and less data leaves. An essay agent that suddenly goes online or draws pictures can drag in junk. The narrower the access, the more predictable the work.

How do you know the agent is ready?

Run 3 tests, one of them tricky. Normal requests show it does its job, and the tricky one shows it sticks to the rules, like not translating everything for you.

Skills: your own agent in 10 minutes

Lesson 5 of 22 in the course “AI Agents for Life”. Free, with animations and a quiz.

You can pack a brief

A good brief is a pain to rewrite every time. So all three big assistants have a way to “pack” it into your own mini-agent. In November 2023 OpenAI launched GPTs: custom versions of ChatGPT you build without code. In March 2025 Google opened Gems in Gemini to everyone for free. In October 2025 Anthropic released Skills for Claude: a folder with instructions and files that Claude loads when it’s needed. That December, Skills became an open standard. Different names, same idea: set it up once, then just use it.

What it’s made of

Any “agent of your own” has four parts. Instructions are your brief: role, rules, stop rules. Knowledge is the files it relies on: notes, a timetable, a list of favourite recipes. Tools are what it’s allowed to do: search the web, make images, crunch tables. And tests: three requests you use to check it works right. That’s how OpenAI itself described GPTs at launch: instructions, extra knowledge and a set of skills. The fourth part, tests, is ours: without it you won’t know when the agent breaks.

When it’s worth making one

Not every request deserves its own agent. A simple rule of three. You wrote a brief once: that’s just a chat. You repeated it: save the text in your notes. You’re writing it a third time: the task is regular, so it’s time to pack it. That’s how Liza ended up with three mini-agents: an essay checker for her uni’s requirements, an English tutor and a “receipt sorter” that puts spending into categories. Each took 10 minutes to build and saves her half an hour a week.

Build it in 10 minutes

The steps are the same everywhere. You create a Gem at gemini.google.com, a GPT at chatgpt.com/create, and in Claude it’s easiest to start with a project. Step one: paste your brief into the instructions field. Two: add files, but only ones you wouldn’t mind showing. Three: turn on only the tools you actually need. Fewer buttons mean fewer ways to go wrong and less data leaving. Four: run three tests, one of them tricky, and fix the instructions wherever the agent slipped.

A public agent is an open book

You can share your GPT with a link, which is handy. But in 2025 researchers tested 10,000 real public GPTs: over 98.8% gave away their instructions in response to tricky prompts. Back in 2023, journalists reported that files uploaded to some GPTs could be downloaded. The takeaway is simple: anything in a public agent’s instructions and knowledge, treat as published. Your study notes, sure. Your passport, bank statements, other people’s contacts, never.

Try it: a tutor

Build your first mini-agent. Open Gemini → Gems → new Gem, or chatgpt.com/create, or a new Claude project, and paste the brief below into the instructions. Change the level and language to suit you. Then three tests: “topic: café,” “translate the whole dialogue for me” (it should give a hint, not a translation) and “what did I study last time?” If the agent makes up an answer to the last one, add to the instructions: “if you don’t remember, say so honestly.” Liza does 15 minutes with a tutor like this on the bus.

Check yourself

What should she do?
Correct answer: Pack the brief into a Gem, a GPT or a project. A third repeat is the signal that the task is regular. Your own agent keeps the whole brief, so nothing gets lost.
What would you tell him?
Correct answer: Nothing personal in a public agent: files and instructions can leak. A 2025 study: over 98.8% of public GPTs gave up their instructions to tricky prompts. A passport doesn’t belong there.
Why is it better to keep only the tools you need?
Correct answer: Fewer ways to go wrong and less data leaves. An essay agent that suddenly goes online or draws pictures can drag in junk. The narrower the access, the more predictable the work.
How do you know the agent is ready?
Correct answer: Run 3 tests, one of them tricky. Normal requests show it does its job, and the tricky one shows it sticks to the rules, like not translating everything for you.

← Memory: an agent that knows you Deep research: the researcher agent →