Claude Code says the task is done, but is it? Here’s how to use Claude Code the way Anthropic’s Claude Code team checks its work: a Claude Code setup with Claude Code skills, hooks and a review loop that make Claude AI prove each feature works, plus Claude Code tips for what still slips through.
Try Unstoppable at app.unstoppable.ai/code?utm_source=ai_labs to run Claude Code, Codex, and Cursor side by side, with 25 free AI credits every day.
Community with All Resources:
The Roundup, our daily newsletter covering the AI stories that matter. Join now:
Got a software idea but don’t have the team to build it? Autometa turns your ideas into production-ready software. Check out to see what we can build for you.
Article link:
AI agents write a lot of code, call the task done, and miss things you asked for. Anthropic’s Claude Code team shared how they check their agents’ work: more than twice as many changes, and the share they had to undo dropped by over 40%. We turned it into a workflow you can follow.
The workflow
– Make Claude prove it’s done: the write-evals skill turns each feature into evals and a design mockup, and hooks block building until you’ve reviewed them.
– A separate reviewer: the full review agent is Team and Enterprise only. The free Claude Code option is the built-in code review command.
– How to set up Claude Code to follow your review rules: put them in REVIEW.md and link it from CLAUDE.md.
– AI automation with a review loop: builder, verifier and reviewer agents run in rounds until every eval passes.
– Fix it at the source: repeat mistakes become one-line rules in REVIEW.md.
– Match the check to the risk: rate each feature low, medium or high before you build it.
– Stop the loop from running forever: our loop stops after five rounds, or early when fixes stop working.
– When something still gets through: a first-time-user agent tests the live app, and a fix-a-bug skill finds and fixes the cause.
– Safe Claude Code editing with Git: a branch for each feature, so a bad change rolls back.
Who this is for
– Vibe coding without reading the code? This is how to use Claude Code so Claude AI checks its work before your users do.
– A next step after any Claude Code for beginners tutorial, built for your own Claude Code projects.
– Not a Claude Code full course, one complete Claude Code setup.
– Codex vs Claude Code? The ideas carry over to any AI agent, Codex or ChatGPT included. The skills we built are for Claude Code.
– The Claude Code skills from this video are in AI Labs Pro.
0:00 Intro
0:45 Why Review Matters
2:41 Evals Skill
5:42 Separate Reviewer
6:27 Review.md
7:21 Review Loop
8:54 Sponsor
9:55 Learned Rules
10:41 Risk Levels
11:42 Stopping the Loop
13:19 First-Time User Test
14:49 Fix a Bug Skill
Hashtags:
#ai #Claude #ClaudeCode #ChatGPT #Codex #ClaudeAI #aiAutomation #vibeCoding
コメント