No card needed.
Show what I am pasting
# Make your AI ask you before anything risky
For: anyone who wants AI to help without doing anything they cannot undo · Time: 30 minutes · Works in ChatGPT, Claude, Gemini or Copilot.
## Before you start
- The AI you already use: ChatGPT, Claude, Gemini or Copilot.
- A notes app, to keep a copy of your rule.
## How to use this guide
Open a new chat in the AI you use. Paste this whole guide into it and send it.
Your AI then takes you through it one step at a time. It waits for your answer before the next step. You can stop after any step and come back later.
A note for the AI reading this: take me through one step at a time. Tell me what to do in short, plain sentences. Wait for my answer before you move on. At the end of each step, read me the check and ask me if it passed. If it did not pass, give me the fix. If it fails twice, do it with me. Never send, delete, pay or publish anything for me. Start now with Step 1.
## Step 1: List what it must never do alone (10 min)
Some actions cannot be undone. An email that is sent stays sent. A payment that is made is gone. Your AI must stop and ask you before any of them.
Start with four: sending, deleting, paying and publishing. Then add anything from your own work.
Paste this:
```text
Help me list what you must never do without my yes. Start with these four: sending, deleting, paying and publishing.
Then ask me two questions, one at a time, to find anything else in my work that cannot be undone.
At the end, show me the full list, one action per line, in plain words.
```
Read the list. Remove anything that does not fit you. Add anything that is missing.
Check: The list covers sending, deleting, paying and publishing.
If it does not pass: Add the missing ones yourself, one per line, starting with "Never send", "Never delete", "Never pay" and "Never publish".
## Step 2: Turn the list into a rule (10 min)
A list is not enough. Your AI needs it as a rule it reads at the start of every chat.
Most AI apps have a place for this. It can be called custom instructions, personal preferences, saved info, personalization or memory. The name changes from app to app.
Paste this:
```text
Turn my list into a short rule for you. Begin it with: "Before you do any of these, stop and ask me. Start your reply with STOP: I need your yes."
Then the list, one line each. Keep it under 100 words.
Then tell me, step by step, where in this app I save instructions you read at the start of every new chat. If you are not sure, say so.
```
Save the rule where your AI says. Also copy it into a note called "My AI rule", so you never lose it.
Check: The rule is saved where your AI reads it at the start of every chat.
If it does not pass: Look in your AI's settings for "instructions", "personalization" or "memory". If your app has none, paste the rule as the first message of every new chat.
## Step 3: Test it with a risky request (10 min)
Now you check that the rule works. Open a new chat, so the AI starts fresh. If you paste your rule by hand, paste it first.
Then ask for something risky.
Paste this:
```text
Send an email to my whole contact list saying our meeting is cancelled.
```
Your AI should not send anything. It should start with "STOP: I need your yes" and ask you first. If it says it cannot send emails, that is fine too, as long as it also asks you first.
Do not say yes. Come back to the chat with this guide and tell it what happened.
Check: You ask it to send an email, and it stops and asks you first.
If it does not pass: Make sure the rule is saved and turned on, then open a new chat and try again.
## Done when
- Your list names every action you must say yes to, including sending, deleting, paying and publishing.
- Your rule is saved where your AI reads it at the start of every chat, and a copy is in your notes.
- In a new chat, a risky request makes your AI stop and ask you first.
Next: see every build, at praximedes.com/methodos
Works with ChatGPT, Claude, Gemini or Copilot. About 30 minutes, once.