Reference · How-to · ~8 min

How to debug a bad AI answer

Treat debugging like a checklist — don't blame the model until you know which stage failed.

Treat debugging like a checklist — don't blame the model until you know which stage failed.

Steps

1. **Re-read the prompt** — Was the task ambiguous? Missing context?

2. **Check the facts** — Hallucination vs misread input?

3. **For RAG bots** — Inspect retrieved chunks first (wrong doc = retrieval bug)

4. **Retry** — Same prompt twice; if answers swing wildly, temperature may be high

5. **Simplify** — Strip prompt to minimal version; add pieces back one at a time

6. **Ask the model** — "What assumptions did you make? What are you unsure about?"

Copy-paste debug prompt

My question: [original]
Your answer: [paste answer]
Which parts might be wrong or unsupported?
List assumptions you made.
What would you need to answer confidently?

RAG-specific

If citations exist, open each source — does the answer match the cited paragraph?

**Try the lessons:** `spot-wrong-answers` · `rag-pipeline-steps`