Create a short review checklist for [specific AI task/output].
Use these source rules and likely failure modes. For each check, state
what to inspect, the independent evidence, an example of failure, the
stop or correction action and what would count as a checked result.
Prioritise checks that affect this output's use. Do not add generic
confidence scores or declare quality from the assistant's own assurance.
Keep advanced checks optional where they are not needed for the first job.

---

Apply the checklist to this labelled flawed example.
For each finding, point to the source and affected output detail.
If a known material fault is not covered, add or sharpen the relevant
check. Remove checks that cannot be evaluated from the available evidence.
Explain the correction request rather than giving the person a score.

---

Check [name] failed because [output detail] differs from [source evidence].
Correct that specific detail while preserving [sound parts].
Return the revised output and identify the changed section or field.
Repeat the affected check and any related boundary check.
Leave unresolved evidence as a focused question; do not mark it passed.
