Examples
Before and after, with the reasoning in between.
Each example shows the original request, a plausible first draft, what the critique caught, and the final prompt you would hand to Claude Code or Codex.
Reproduced from the canonical examples file at commit d0babf8f (line breaks reflowed). These are authored illustrations, not evaluation results; see the evaluation protocol for measured runs.
Small bug fix
Write a Codex task. The Save button shows success even when the API returns an error. Preserve the API shape. Don't deploy.
Fix saving and run tests.
Lost the visible symptom, API compatibility and deployment boundary. No check demonstrates that an error stops the false success state.
Fix Save so an unsuccessful API response is not reported as success. Preserve the API shape. Locate the existing save path and verify the requested failure behavior. Do not deploy.
Optional suggestion: use the existing error-display pattern if one is available.Authorized implementation
Write a Claude Code handoff: add CSV export to the existing report view; reuse current columns, no backend change. Edits, tests and an MR are authorized; merge and deployment aren't.
Implement CSV export, then ship it.
"Ship" invents merge/deploy authority. Missing column and backend constraints.
Add CSV export to the existing report view using its current columns. Make no backend changes. Edits, tests and an MR are authorized; merge and deployment are not. Verify the export represents the chosen columns as valid CSV.
Optional suggestion: consult nearby export patterns; retain existing empty-result conventions if applicable rather than choosing a new one in this handoff.Performance with measured acceptance
Write a Codex task: reduce our existing 10,000-row filter benchmark's p95 from 900 ms to under 300 ms on the same runner, retaining exact results. No new dependencies; no deployment.
Optimize the filter with caching.
Prescribed an unchosen mechanism and dropped measurement conditions.
Reduce p95 latency in the existing 10,000-row filter benchmark to below 300 ms on the same runner and benchmark protocol used for the 900 ms baseline. Preserve exact filtering results and add no dependencies. Measure the stated benchmark under its existing protocol to verify the target and unchanged results. Do not deploy. The optimization mechanism is not prescribed.When no rewrite is needed
If the original task already states the goal, constraints, authority and a feasible check, PromptingCraft.com says so and keeps your text exactly:
“The original prompt is already suitable. No material revision is necessary.”
No-change still goes through the contract and source checks.