Cette page n’est pas encore traduite.

Use your own AI to write

If you already use an AI app, it can help you write instructions with your own subscription. This page shows how Claude or ChatGPT works on your instructions from a chat, how your browser's built-in AI works in the editor, and how their changes reach your instruction.

None of these use your assistant allowance.

Before you start

  • Account: for Claude or ChatGPT, you need an instruction.md account and an account with the AI app. Your browser's AI also works on instructions kept in this browser without an account; sharing and publishing need one.
  • Role: the AI acts as you. It can do only what your role on the instruction lets you do (see People and roles).
  • Plan: connecting an AI app, and using your browser's AI, work on every plan.

Write with Claude or ChatGPT

First connect the app to instruction.md, once. Connect Claude or ChatGPT shows how. When you connect, you sign in, so the app works as you.

Then ask for what you want in the chat, in plain words. For example:

  • "Write an instruction for our support agent."
  • "Tighten the rules of my refunds instruction."
  • "Check my coding-agent instruction for problems and fix them."

How its changes reach your instruction

The AI app changes the draft, never a published version. What happens to a change depends on your role:

  • If you can only suggest changes (Contributor), every change it makes waits as a suggestion. A maintainer accepts or rejects it.
  • If you can edit (Maintainer or Admin), it changes the draft directly. Ask it to "suggest" if you want to review each change first. Then its changes wait as suggestions you accept or reject.

Nothing reaches the agents that follow the instruction until a new version is published. The AI app can't publish on its own: it shows you what publishing would do, and you confirm.

The views inside the chat

Where the app shows the views of the apps you connect, instruction.md draws your instruction in the conversation. Claude and ChatGPT can both show them.

The draft view inside Claude.

The draft. When the AI opens or creates an instruction, you see:

  • its name, and what you may do with it, such as you can edit it or you can suggest changes;
  • its state, such as Draft, not published yet or v3 is live · the draft has changes;
  • the text, drawn as the site draws an instruction. Show Markdown shows the plain text, Show the document switches back;
  • the changes waiting, each with Accept and Reject, and Accept all and Reject all. If you can't accept them, it says A maintainer of this instruction accepts or rejects these.;
  • what the editor's checks find, such as No problems found.;
  • a few ready requests, such as Check it for problems and fix them or Make the rules clearer and testable, and a box for your own: Ask for a change…. A request goes to the conversation, as if you typed it.

You can also work on it yourself, there in the chat:

  • Edit (or Suggest an edit) lets you type in the text. Save saves it to the draft. A contributor's button is Send as suggestions.
  • Comments lists the open comments. You can reply, resolve one, or add your own.
  • Publish asks What changed, then shows what will happen. It publishes only when you confirm.
  • Full screen makes the view larger, where the app allows it. Open in the editor opens the instruction in the full editor.

The view follows the draft while it is on screen, so each change the AI makes shows up without reloading.

Your instructions and their history. When the AI lists your instructions, their versions, or the changes someone offered, you see a list to browse. You can open an offered change and review it there.

Write with your browser's AI

Some browsers have an AI that can use a page's own tools, such as Chrome (as a trial) and the browser in the ChatGPT app. In such a browser, the AI can work in the editor directly, as you, with what your role lets you do. It can:

  • read the open instruction, check it for problems, and suggest changes. If you are a Maintainer or Admin and ask it to, it edits the draft itself;
  • list, accept and reject suggestions;
  • read, write and resolve comments;
  • list your instructions and folders, and open, create, rename, copy and move them;
  • read an earlier version, and put one back in the draft;
  • publish a version, make an earlier version the live one, or revoke a version, if your role lets you;
  • see who an instruction is shared with, its read links, and your workspace's people and settings;
  • run the editor's own commands.

Open the instruction in the editor first. Then ask your browser's AI what to do.

It asks you first

Anything that leaves the draft, or can't be undone, is put to you in the editor first. The question says Your browser's agent asks for this. It happens only if you agree. For example:

  • publishing a version: Publish v4 of "…"?
  • putting an earlier version or checkpoint back in the draft;
  • sharing with someone, taking their access away, or changing who can read the instruction;
  • making or revoking a read link;
  • approving or declining someone's request for access;
  • moving an instruction or a folder to the trash, archiving, and some of the editor's commands.

Making an earlier version the live one, or revoking a version, asks you the editor's own question, the same as when you do it yourself.

The editor asks you before the browser's AI publishes.

Press the button that says what will happen, such as Move to trash or Make the link, to agree. Press Cancel to refuse. The AI is told you said no, and asked not to ask again unless you tell it to.

A page your browser's AI read elsewhere could try to tell it to share an instruction. That is why only you can agree to these steps.

Some things the browser's AI can never do, such as signing you out or turning on developer mode.

Edit with Claude, with your own key

Which to choose

The assistant Claude or ChatGPT Your browser's AI
Where you work In the editor In the chat In the editor
Accounts you need instruction.md instruction.md and the AI app A browser with a built-in AI; instruction.md to share or publish
What it costs Your plan's assistant credits Your AI app's subscription Whatever your browser's AI costs you
What it can change The open instruction Any instruction your role allows The open instruction, and your instructions and folders
Can it publish? No, it opens Publish for you Only after you confirm Only after you agree in the editor
Where your words go OpenAI, in the United States The AI app's company Your browser's AI

Check it worked

  • In the editor, open the instruction. Suggestions from the AI are in the text, waiting for review.
  • Open History to see each change to the draft and who made it.
  • On the Instructions page, an instruction the AI created is listed with the others.

Undo or stop

  • Reject a suggestion you don't want, or Reject all.
  • To undo direct changes to the draft, restore an earlier checkpoint from History.
  • To stop an AI app from working on your instructions, disconnect it in Your account ▸ Connected apps (see Connected apps).

If it doesn't work

What you see What to do
The chat shows text but no view of the instruction Your AI app doesn't show the views of the apps you connect, or not on this device. Ask it to open the instruction in the editor instead.
This app can't send a message here: type it in the conversation instead. Type the request in the chat.
This app can't change its size here. The AI app doesn't allow a full-screen view here. Use Open in the editor.
Fix the problems that would stop publishing first Ask the AI to fix them, or fix them in the editor. See Check and fix.
Your browser's AI can't find the editor's tools Open the instruction in the editor first. Only some browsers let their AI use a page's tools.

Questions

Does my AI app see all my instructions?

It sees what you can see, because it acts as you. It reads an instruction when it needs to answer you. Disconnect it in Your account ▸ Connected apps to stop it.

Can my AI app publish without me?

No. Publishing asks you first, with what will happen. It goes ahead only when you confirm.

Does this use my assistant allowance?

No. Claude, ChatGPT and your browser's AI run on their own accounts. Only instruction.md's own assistant uses the allowance.

Who sees what I write in the chat?

The company behind your AI app, under its own terms. instruction.md receives only what the app sends to work on your instructions.

For developers: how an AI app connects and the views it draws.

Last checked against the product:

Was this page helpful? Report a problem