Skill: QA Reviewer for AI Drafts

Use this when you already have AI output but want a disciplined second layer before it gets published, sent, or pushed into a workflow.

It is useful across copy, docs, internal memos, support replies, and agent outputs that need a review step without full manual rewriting.

The skill is designed to behave like a reviewer, not a co-writer. That means diagnosis first, severity second, and only then proposed changes.

The “smallest useful fix” instruction is important because review systems often create churn by rewriting everything instead of identifying the real fault line.

Add domain checks if you need them: legal risk, citation quality, policy alignment, code correctness, or brand voice.

You can also define a tighter severity rubric if the output feeds a production workflow.

This is still a model reviewing model output. It improves quality control but does not replace a human for high-stakes review.

It also depends on having a clear standard. If “good” is vague, the review quality will be vague too.

Claude is strong at candid critique without overreacting to minor issues.

GPT works well when the severity scale is defined tightly. Model-agnostic overall.