/015-make-ai-grade-its-own-work

Reviews Claude's own output against a fixed four-question checklist (completeness, tone consistency, uncertain facts, generic or vague spots) and returns a short list of exactly what needs fixing. Use right after Claude produces writing, research, or any output the user is about to use for something real, or when the user asks to "review", "check", "grade" or "double check" what Claude just produced, or says "run the review" or "is this ready to send?".

Self-Review When to use

Run this right after any piece of writing, research, or output the user is about to use for something real. Skip it for quick throwaway answers.

Target

Review what Claude just produced in this conversation. If it is unclear which output the user means, ask in one short line before reviewing.

Checklist

Apply the same four questions in every review, whatever the original task was.

Completeness: Does it actually answer what the user asked, with nothing missing? Tone: Is the tone consistent from start to finish? Certainty: Are there any claims, numbers, or facts Claude is not fully certain are accurate? If so, flag exactly which ones. Specificity: Is there anything generic or vague that a specific detail would improve? Output

Give a short list of exactly what needs fixing, and nothing else.

One line per issue, pointing to the exact sentence, claim, or passage. For uncertain facts, name the specific claim or number. No preamble, no praise, no recap of the checklist, no rewritten version. Leave out any question that found nothing. If nothing needs fixing, say so in one line.

After the list, stop. Apply the fixes only if the user asks.