Every week, you do the same thing with your assistant: paste the same context, remind it of the same format, correct the same slip-ups.

A tailor-made assistant solves that. You set it up once for one specific task, and it's ready every time you come back. In ChatGPT, they're called custom GPTs; in Gemini, Gems; in Claude, Projects play a similar role. The name, location and features vary by tool and change with updates, and some options may require a paid plan depending on your subscription: check your tool's official help pages. The method stays the same everywhere.

Custom instructions, dedicated assistant, agent: three different things

Custom instructions set up your assistant for all your conversations: language, tone, honesty about uncertainty. A dedicated assistant is specialized: it does only one thing, but always the same way.

An AI agent goes further still: it receives a goal and chains actions on its own using tools (search, read, send). The dedicated assistant described here stays in the conversation: it answers you, and you decide what to do with the result. That's exactly what makes it a good starting point.

Pick a single task

The first mistake is wanting an assistant that does everything. The broader the mission, the more mediocre the answers. Choose a task that ticks three boxes:

  1. It comes up often: at least once a week.
  2. It always follows the same pattern: same inputs, same kind of result.
  3. You can recognize a good result: you can judge in a minute whether it worked.

"Help me at work" is too vague. "Turn my meeting notes into a one-page summary, with decisions and action items" is perfect.

Anatomy of a good dedicated assistant

1. The mission

One or two sentences: for whom, for what, with what result. It's the compass for everything else. If you can't write it simply, your task is still too fuzzy.

2. The instructions

The task's process, step by step: what the assistant receives, what it does with it, what it hands back. Specify the output format, the tone, and what it should do when information is missing: ask a question rather than guess. These are the same building blocks as a good prompt, detailed in how to write a good ChatGPT prompt.

3. The reference files

Most tools let you attach documents the assistant can consult: a style guide, an internal FAQ, a meeting summary template. They're often called "knowledge" or "files." Keep them short and up to date, and say in the instructions when to use them.

4. The examples

One or two examples of the expected result beat a long paragraph. The assistant imitates the structure, length and tone. Choose successful, anonymized examples.

5. The guardrails

What the assistant must never do: invent a figure, date or name that isn't in what you gave it, stray from its mission, answer an off-topic question as if it knew. Also say what to do in those cases: flag what's missing, reply "outside my scope."

An instruction template to copy

Replace the brackets, then cut whatever doesn't serve your task.

Mission: you help [who] to [specific task], in order to get [expected result].

What you receive: [type of input, for example raw meeting notes].

What you do, in order:
1. [step 1]
2. [step 2]
3. [step 3]

Output format: [structure, length, headings or table].
Tone: [register], in English.

Reference files: use [file name] for [specific use]. If the information isn't there, say so.

Example of the expected result:
[paste a successful, anonymized example here]

Guardrails:
- Don't invent any figure, name, date or source that isn't in what I gave you.
- If information is missing or the request is vague, ask a question before answering.
- If the request falls outside your mission, say that it isn't your role.

Test with five requests

Before using it for real, prepare five test requests and note the result of each:

  1. A typical case: the task as it usually comes in.
  2. A hard case: long, messy or ambiguous input.
  3. An incomplete case: a key piece of information is missing. Does the assistant ask a question?
  4. An off-topic case: a request outside the mission. Does it stay in its role?
  5. A trap: a request that pushes it to make things up ("add last year's figures" without providing them). Does it flag what's missing?

Iterate, one instruction at a time

When a test fails, change just one thing: rephrase the instruction in question more concretely, move it higher, or add an example. Then rerun the five tests. If you change everything at once, you won't know what worked.

Keep your five requests: with every change, you can rerun them in two minutes. And a well-tuned assistant can still be confidently wrong: always double-check its figures and names (see why ChatGPT makes things up).

Be careful what you upload

Instructions and reference files are stored by the service you use, and an assistant you share can sometimes reveal its contents to the people using it. So don't upload any sensitive data: personal data about clients or members, health information, passwords, your employer's confidential documents. At work, first check what your organization allows, and review the sharing settings before distributing your assistant.

Going further

Mission, instructions, tools, guardrails: that's the vocabulary of Ovrelia's "Anatomy of an Agent" chapter, which teaches you to tell what answers from what acts, and to set the right limits. It's free and plays in your browser, no account needed: play it now.

To write your instructions block by block (role, context, task, format, constraints, example, tone) and then review them with a checklist, use the Prompt Lab. If your recurring task can run without you, read how to automate your tasks with AI, and find mission ideas in using ChatGPT at work. And to test your reflexes, launch the friend challenge: ten timed questions about AI, and a score to beat.

Frequently asked questions

How do I create a custom GPT?

Pick a specific recurring task, then write its mission, step-by-step instructions, the expected format, one or two examples and guardrails. Add short reference files if needed, then test with five varied requests. Where the feature lives and whether your plan includes it vary: check your tool's help pages.

What's the difference between a custom GPT and custom instructions?

Custom instructions apply to all your conversations and set your general style. A custom GPT, a Gem or a Project is dedicated to one specific task, with its own instructions, examples and files.

Is a custom assistant an AI agent?

Not necessarily. The dedicated assistant described here answers in the conversation and lets you decide. An agent receives a goal and chains actions on its own using tools, which requires stricter guardrails.

Can I put my documents in a custom assistant?

Yes, most tools let you attach reference files. Avoid any document containing personal or health data, passwords or confidential information, and check the sharing settings before making the assistant available to others.