Course Content
Common Prompt Failures and Fixes
Diagnose hallucinations, refusals, and inconsistent outputs — and fix them
The Four Prompt Failure Modes
1. Hallucination
The model generates confident but false information.
Symptoms: Incorrect facts, made-up citations, wrong numbers.
Fixes:
- Add: “Only use information from the provided context. If the answer isn’t there, say ‘I don’t have enough information.’”
- Use RAG to ground answers in real documents
- Ask the model to cite its sources
2. Format Drift
The model ignores your format instructions for complex inputs.
Symptoms: Prose instead of JSON, wrong number of bullets, missing sections.
Fixes:
- Repeat format instructions at the end of the prompt
- Add a few-shot example showing the exact format
- Use schema validation and retry on failure
3. Scope Creep
The model adds unrequested content, caveats, or disclaimers.
Symptoms: Unsolicited warnings, extra sections, hedge phrases.
Fixes:
- “Do not add caveats, disclaimers, or unsolicited advice”
- “Respond with exactly the requested output and nothing else”
4. Ambiguous Interpretation
The model interprets your task differently than you intended.
Symptoms: Off-topic responses, wrong level of detail, wrong audience.
Fixes:
- Add a role: “You are a [expert] writing for [audience]”
- Add a negative constraint: “Do not include [X]”
- Add an example of what you DON’T want
Debugging Workflow
- Identify which failure mode occurred
- Apply the targeted fix
- Test with 5+ diverse inputs
- If still failing, add a few-shot example
