Start with one harmless example
Choose a task where you can tell what went wrong: a missed date, an invented source, a skipped instruction or the wrong format. Remove names, passwords and private records before sharing the prompt and answer with another service.
You do not need to install a new Bot. This is a one-time check you can run in the AI chat you already have. If the same kind of mistake keeps returning, the method can become a reusable skill later.
Ask for one correction
Fill in the four blanks below and paste the request into your AI chat. Ask it for one likely cause and one small change—not a complete rewrite of your setup.
I want to fix one bad AI result. Do not change files, settings, or accounts.
My original task: [paste the exact task]
The answer I got: [paste the failed answer]
What I expected instead: [describe the observable difference]
What information and tools were available: [list only what matters]
State the mistake in one sentence. Separate what you can see from what you are guessing. Name the single most likely cause and point to the task or answer that supports it. Suggest the smallest instruction change that might prevent it. Tell me what remains uncertain. Do not claim the change works until I rerun the same harmless task with only that change.Run the same task again
Keep the original task and sample material. Change only the one instruction you chose, then run it again. Compare the new answer with what you expected. If it improves, keep the change; if not, put the old instruction back.
Ask before changing a standing Bot, a shared skill, access permissions or a scheduled job. A convincing explanation of a mistake is not proof that the proposed fix works.
- Can you point to the exact part of the first answer that failed?
- Did the second run fix that problem without creating another one?
- Did anything get sent, saved or changed without your approval?
Where this idea came from
Adapted from Diagnose My Agent's Mistake on Muse at Work. Bot Cabinet reviewed the public listing but has not run the original Muse workflow or tested these steps in each AI service.