When Claude writes and reviews the code, who checks that it actually works?
Claude Code’s team describes how it checks work largely written and reviewed by AI. Triggered reliability checks, human escalation and protected review instructions aim to replace confident completion claims with evidence, while keeping responsibility for risky changes and release decisions with people.
With roughly 90% of its code written and reviewed by Claude, the Claude Code team needs evidence of completion. Addy Osmani’s November 16 account describes a 39-item reliability checklist, usually triggering three to ten checks per change. A check still failing after three repair rounds goes to a human.
An October team account described one pull request cycling through more than fifty overnight review rounds before its author intervened. Automation can generate work for itself; people still own architectural choices, risk and what ships.
The verification account connects to a concrete protection in Code Review: review instructions come from the base or default branch, never the pull request’s own changes. Teams must protect that branch. The service also differs from the local review command: the former reads REVIEW.md, while the latter relies on CLAUDE.md instead.
Testing can challenge an implementation rather than repeat its author’s assumptions. The linked Hypothesis documentation explains property-based tests: specify a requirement, then generate varied inputs, including edge cases. A sorting test can check both retained values and ordering. Like inspecting a lift under different conditions, this asks what would break the claim, not merely whether one demonstration worked.
Official hook documentation shows how failing tests can keep a task incomplete and return feedback to the model. Stop-time checks can likewise request more work. Completion becomes an executable condition rather than a confident sentence. People then decide which failures block release and where those conditions still leave blind spots.
Share this article
What do you think of this story?
Anonymous feedback, up to 5 reactions per article.
Reader comments
Comment without signing in, or sign in to display your member name.
Discuss the technology and the story. Hate, politics and war, threats, abuse, personal accusations, sexual content, ads, links and personal information are not allowed.
No comments yet. Be the first to share.