Diagnose an unsatisfactory AI answer before rewriting the prompt. The cause might be an unclear task, missing evidence, a formatting failure or an unavailable tool. Match the repair to the observed defect.

Use a symptom-to-check table

SymptomPossible causeRepairAcceptance check
Answer solves the wrong taskDeliverable was ambiguousName one artifact and its purposeOutput serves the stated task
Summary invents a dateDate absent or confused with another sourceSupply labelled evidence; keep missing values unknownEach date matches a passage
JSON cannot be parsedFormat instruction failedValidate with a parser; use a supported schema feature if building an integrationParsing, fields and factual values all pass
Claims to have booked a meetingNo calendar action or confirmationUse the authorized calendar workflowActual event confirmation exists
Current information is staleNo relevant retrievalRetrieve a current primary sourceOpened source supports the dated claim

OpenAI’s structured-output guide explains schema constraints and remaining value errors. Its web-search documentation covers retrieval. These are different tools for different failure classes.

Example: a missing booking link

Source: the workshop is Saturday at 10:00 in Room 4. Bring a laptop. Booking is required, but the link is pending. Candidate: “Book at library-workshop.example to guarantee your place.”

The draft invented both a URL and a guarantee. More enthusiastic wording or a longer persona instruction would not supply the missing evidence.

Correct the invitation using only the source facts.
Remove the invented URL and guarantee.
Keep booking required and state that the link is pending.
Preserve Saturday at 10:00, Room 4 and the laptop requirement.
List the detail the organizer must confirm before publication.

Example corrected artifact

Join the workshop on Saturday at 10:00 in Room 4. Bring a laptop. Booking is required; the booking link is pending.

Organizer question: Which confirmed booking link should we publish?

Keep a useful failure record

Save the source, faulty output, cause, repair and acceptance result. Try a fresh input to check whether the repair generalizes. A short prompt can be adequate when it contains the information the task needs; length alone is not a diagnosis.

For practice, classify one failure using the table, apply one repair and inspect the full result. Use a bounded follow-up for a local defect, or review gates when a bad intermediate artifact would affect later stages.