How do I debug a failing prompt?
Reduce variables: shorten context, simplify the ask, and specify the output. Reproduce the failure on a tiny example set. Change one thing at a time—role, examples, or constraints—and note the impact. If it still fails, try a different model or add a staging step. Logging and a small eval harness make debugging fast and repeatable.
In practice
Ask the model to explain how it interpreted your prompt—it often reveals the mismatch.
Where to start
- [01]Create a 10-case failure set.
- [02]Strip prompt to essentials and rebuild in steps.
- [03]Swap models to isolate behavior vs. content.
- [04]Document the fix and add a regression test.
debugging promptsfix bad AI outputprompt troubleshooting