Models make mistakes, but they're often good at spotting mistakes when asked to check. Building in a review step improves reliability.
Simple Self-Checks
At the end of a prompt: "Before finalising, check that every figure matches the source and every requirement in the brief is met. Fix any problems."
Separate Review Passes
For important work, run a second prompt that reviews the first output against explicit criteria:
- Are all claims supported by the source?
- Is anything required missing?
- Are there calculation errors?
- Does it follow the format?
Then ask for a corrected version.
Checklists
Give the reviewer a checklist derived from your requirements. Specific checks beat "review this for errors".
Different Perspectives
Ask the model to critique as a sceptical reader, a subject expert or the intended audience.
Verify With Tools
For facts and numbers, prefer deterministic checks: run code, recalculate totals, validate JSON against a schema, check quotes against the source.
Limits
A model may confidently approve its own errors, especially factual ones it believes. Self-checking reduces errors; it doesn't eliminate them. Keep human review for high-stakes outputs.