What exactly went wrong?
Name the behavior: an invented claim, a rejected approach, a missed format, or a commitment you did not approve. Ask AI to propose one narrow instruction that could reduce that specific failure.
Should that become a standing rule?
Review the proposal before adding it to the project instructions. Check whether it applies to future work and whether it could interfere with something else. A one-time exception belongs in that task, not in permanent guidance.
Did it change the next result?
Confirm the approved instruction is actually saved, then check the next relevant output. A correction in a chat is not the same as an edited standing instruction. Saving the rule is not proof it will be followed. Keep it only if it helps, and narrow or remove it when it becomes outdated or creates another problem.
Sources
Sources
- Jeff Fargo, Fargo Talks: Why ChatGPT Repeats Mistakes
- OpenAI Help Center: Projects in ChatGPT
- OpenAI Academy: Using Projects in ChatGPT
- OpenAI Help Center: Prompt Engineering Best Practices for ChatGPT
Transcript
Verbatim from the Fargo Walks clip, as recorded.
When AI messes up, don't correct it. You want to ask it? What rule can you create to prevent this from happening?? Take that rule and put it in your project instructions. Then you're fixing the system instead of wasting your time going back and forth in that chat. I'm Jeff Fargo. I can help you streamline and scale with AI. Go forth.
