AI output often looks finished before it is. A confident paragraph can still miss half the question or include a date nobody checked. This skill makes the assistant re-read its own work as a skeptical editor before you rely on it.
What it does
Run it right after an answer you're about to send, publish or act on. It checks the output against the same five questions every time:
- Did it answer the ask? Anything missing or answered to a different question.
- What's uncertain? Every fact, number, date, name, link or capability claim it can't fully stand behind.
- Where is it vague? Generic lines a specific detail would improve.
- Is it consistent? Contradictions, shifting tone, numbers that don't add up.
- Would it hold up in use? The most likely thing to go wrong if you use it as is.
What you get
Only a list of what to fix, one line per issue, each pointing at the exact sentence. No praise, no recap and no rewritten version. If nothing needs fixing, it says so in one line.
It fixes things only when you ask, and when it does, it checks uncertain facts rather than just rewording them.
Try it
- "Review that"
- "Grade your last answer"
- "Is this ready to send?"
Adapted from the Make AI Grade Its Own Work idea in AI Skill of the Week, rewritten from scratch.
Install
In Claude Code, copy the folder into your skills directory:
git clone https://github.com/ecemkaraman/ai-skills.git
cp -r ai-skills/skills/self-review ~/.claude/skills/
In the Claude app, zip the self-review folder, then go to Customize → Skills, click +, choose Create skill → Upload a skill, and upload the zip. Code execution needs to be on in Settings → Capabilities.