Adjusting Approach Based on Feedback and Results
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.
Iterate deliberately (design guidance)
- Define "better" before you change anything: a checklist, a sample answer, or a score (for example three criteria rated 1 to 5). Anthropic's own method for prompts is the same: define success criteria, test on cases, refine, test again (verified).
- Change one thing at a time: add an example, then tighten the format, then adjust tone. Re-run the same test inputs.
- Keep a small set of test inputs (three to five, including a messy one) and re-run them after each change.
- Compare outputs side by side against your criteria.
- Record versions: which prompt version, what changed, what improved.
- Use real feedback: from readers and users, from reviewers' corrections, from errors found later.
- Know when to stop: when it meets the criteria reliably, lock the version.
Using feedback in the conversation
- Be specific: "The second paragraph overstates the savings; use the 12% figure and say 'estimated'" works far better than "make it better".
- Tell Claude what to keep as well as what to change.
- Provide a corrected example and ask it to follow that pattern.
- If a conversation has drifted, restart with an improved prompt instead of patching (D3.1).
Changing the approach
If several tweaks do not help:
| Signal | Change of approach |
|---|---|
| The task is too large | Split it into steps (D1.2) |
| Missing information | Add source documents, a Project, or a connector (D5.3) |
| Needs current facts | Use Research with web search |
| Needs more depth | Try a more capable model (D3.2) |
| Needs a person | Escalate to an expert (D2.1) |
Learn from results over time
Track simple measures: time to a usable result, how often output needs major correction, errors found after review, and user satisfaction. Review them monthly and update templates (D7.3).
Exam traps
- Judging "better" by feel with no criteria.
- Changing everything at once.
- Vague feedback ("not quite right").
- Never recording what worked.
- Endless tweaking of a conversation that should be restarted.
References
- Define success criteria and build evaluations — checked Oct 16, 2026
Check your understanding
Why change one thing at a time?Show answerHide answer
Answer
So you can tell which change caused the improvement.What makes feedback to Claude effective?Show answerHide answer
Answer
It is specific and says what to keep and what to change.When should you restart instead of tweaking?Show answerHide answer
Answer
When the chat has drifted or grown long and the improved prompt can carry the context.