Prompt repair
Fix the failure, not the entire prompt.
Improve a weak ChatGPT or Claude answer by identifying the failure, supplying missing context and making a focused revision you can verify.
Updated
Describe what went wrong observably
‘This is bad’ gives little direction. ‘It omitted the implementation cost’ identifies a repair. Separate failures of factual grounding, scope, reasoning, voice and output format. More effort is not a remedy for a missing source, and more creativity is not a remedy for the wrong audience.
Save the original brief and result. Comparing changes is easier when you can see what improved and what was lost.
| Observed failure | Useful repair | What to preserve |
|---|---|---|
| Generic argument | Add audience evidence and a concrete difference | Supported claims |
| Wrong conclusion | Inspect assumptions and source interpretation | Correct calculations |
| Wrong voice | Supply an approved example and observable rules | Meaning and exact facts |
| Unusable format | Define the reader's next action and output structure | Essential qualifications |
Supply the missing input before requesting a rewrite
If the assistant cannot know your house style, attach a sample. If a report lacks an appendix, provide it or narrow the question. Ask what information would materially change the answer, not a long questionnaire about details that do not matter.
Compare this output with my original goal. Identify the three most consequential gaps and classify each as missing input, unsupported claim, interpretation error or presentation issue. Ask only for information that would change the repair. Do not rewrite yet.
Change one important dimension
A full rewrite may discard the useful structure or introduce new claims. State the change and the parts to retain. For an image edit, identify the area and details that must not drift. For analysis, keep verified facts fixed while examining the disputed inference.
Revise only [specific weakness]. Preserve [correct facts, structure or approved wording]. Use [new input]. Explain any change that affects meaning, and flag requirements that still cannot be met. Do not add unrelated improvements.
Use a second case before saving the prompt
A prompt can work on one friendly example and fail on a difficult input. Try a second case with a missing value, conflicting source or different audience. Record what you checked rather than calling the prompt universally tested.
If comparing ChatGPT and Claude, keep the task, inputs and assessment criteria equivalent. Record the actual model, settings and date. A small comparison can teach you about your workflow without establishing a general model ranking.
Evaluate this revised brief on [second input]. Check the same requirements as before. Identify where the brief depends on assumptions unique to the first case. Suggest the smallest reusable change and keep decisions requiring human judgement explicit.
Common questions
Should I start a new chat when the answer is poor?
A focused correction may be enough. Start fresh when the accumulated context is misleading, and carry across the verified brief and source roles.
How do I know the prompt improved?
Compare results against the same observable requirements and check a second representative input. Record remaining failures as well as successes.
Keep exploring
About this workflow
This is an original AI-assisted editorial workflow. Illustrative scenarios are not customer case studies, and the prompts have not been systematically evaluated on live models. Check consequential outputs against your sources and use qualified review where needed.