A bot gets something wrong. You correct it. It says, "You're right, here's the fixed version." Problem solved?
For today, yes. Tomorrow it'll make the same mistake.
Corrective action in ISO
ISO covers nonconformity and corrective action. When something goes wrong, you:
- Fix the problem in front of you. Contain it.
- Find the cause. Why did it happen?
- Fix the cause. So it doesn't happen again.
- Check that the fix worked.
- Update the system if needed.
The point is step 3. Fixing the part isn't enough. You fix the process that made the bad part.
The bot version
For a bot, the "process" is mostly its instructions: its principles, playbooks and skills. When a bot makes a mistake, the cause is often one of these:
| Cause | Example | Fix |
|---|---|---|
| Missing rule | It didn't know to cc the project lead | Add the rule to the skill |
| Unclear rule | "Be brief" meant different things | Rewrite it specifically |
| Conflicting rules | Two instructions disagree | Pick one, delete the other |
| Missing source | It guessed a price | Point it to the price file |
| Rule only in chat | You said it once, it got lost | Move it into the instructions |
The best question a bot can ask
A good bot does this on its own. When you correct it, it asks some version of: "Should this be the new default?"
That's the bot running corrective action. It's asking whether your fix belongs in its standing instructions or just in today's answer. If you say yes, it updates the instruction file. If you say no, it was a one-off.
Check that it worked
After the fix, test it. Give the bot the same kind of task and see if the mistake comes back. If you have a tester bot, add a check to its script so the mistake gets caught if it ever returns. See Your tester bot is your internal auditor.
A real example
Our AN Creation tester skill is on revision 8 as of October 20, 2026. Each revision came from finding something that didn't work and fixing the instruction, not just the one run. The skill files carry revision numbers so it's clear which version a bot is working from.
Don't over-correct
Not every mistake needs a new rule. If a bot's instructions grow a new rule for every slip, they get long and start to conflict. Ask: is this a pattern, or a one-off? Patterns get fixed in the system. One-offs get fixed in the answer.
What this means for you
- When you correct a bot, ask yourself: will this happen again?
- If yes, fix the instruction, not just the output.
- Retest after the fix.
- Keep a short note of what changed and why. See Change on purpose.
Fixing answers keeps you busy. Fixing causes makes the bot better.