Diagnosing Underperforming Prompts and Weak Output
Audience: professionals who use Claude at work, no coding or API knowledge needed. Verified = checked against Anthropic's documentation or help center on 2026-10-16. General background = standard regulatory or industry knowledge, not an Anthropic statement. Design guidance = widely taught practice.
Diagnose by symptom (design guidance, built on verified prompting principles)
| Symptom | Likely cause | Try this |
|---|---|---|
| Too vague or generic | The ask was vague; no audience, criteria or examples | State the audience, purpose, format and what "good" looks like; add an example (D1.1) |
| Wrong focus or misses the point | Ambiguous request; competing goals | One task per message; restate the goal; ask Claude to say what it understood before answering |
| Made-up or unsupported details | Facts missing; no permission to say "I don't know" | Provide the source material; allow "I don't know"; ask for quotes (D2.3) |
| Wrong format or length | Not specified | Specify format, length, headings, columns |
| Ignores an instruction | Buried, conflicting or too many instructions | Shorten; put key rules first; remove conflicts; restate in the request |
| Forgot earlier details | Long chat | New chat with a summary, or use a Project (D3.1) |
| Inconsistent between runs | Under-specified task | Add criteria and examples; use a template |
| Shallow reasoning on a hard task | Task too big or model tier too small | Break into steps (D1.2); try a more capable model (D3.2) |
| Tone is off | Audience or voice not given | Give audience and a sample of the voice you want |
| Uses outdated information | No current source | Provide current documents or use Research with web search (D4.3) |
| Hard to verify | No sources or reasoning shown | Ask for sources, quotes and the method |
A diagnosis routine
- Read the output critically and name exactly what is wrong (be specific).
- Check your own prompt first; does it contain what you now wish it had said? Most problems start there.
- Check the inputs: is the material complete, correct and clearly labelled?
- Ask Claude to explain: "What did you understand the task to be? What did you assume?"
- Change one thing, re-run, compare (D7.2).
- If it still fails, change the approach (split the task, add examples, new model, different feature).
Fix the prompt, not just the answer
Do not keep correcting each output by hand if the same weakness will recur. Update the prompt template or Project instructions so the improvement applies next time (D7.3).
Exam traps
- Assuming the model is "bad" before checking the prompt and inputs.
- Changing many things at once, so you cannot tell what worked.
- Adding more instructions when the real issue is missing information.
- Accepting a weak answer and silently fixing it.
References
- Prompting best practices — checked Oct 16, 2026
- Reduce hallucinations — checked Oct 16, 2026
Check your understanding
Output is generic. First fix?Show answerHide answer
Answer
Add audience, purpose, format and a concrete example of good output.Claude invents a statistic. What two changes help?Show answerHide answer
Answer
Provide the source and allow it to say "I don't know" (ask for supporting quotes).What is the first place to look when output disappoints?Show answerHide answer
Answer
Your own prompt and inputs.