ToolNavs Find Useful AI Tools
Submit Sign in
Back to AI information
How the Claude Code Team Verifies Its Work: 90% of Code Written by AI, and What Keeps It from Breaking

How the Claude Code Team Verifies Its Work: 90% of Code Written by AI, and What Keeps It from Breaking

AI information • Admin • • 43 views

On November 16, Anthropic's Claude Code team published its internal approach to verification: roughly 90% of the team's code is written and reviewed by Claude. The trick is not having humans check every line, but a verification loop that lets agents prove their own changes with evidence. The team's point is blunt: without verification loops you can only run one or two Claudes at a time; with them, you can run dozens or hundreds in parallel.

Three kinds of tests — which does the team value most?

The team considers three kinds of tests most worth investing in. End-to-end tests come first: they run real user flows, so a feature either works completely or it fails — they are hard to fool, and they pinpoint exactly which step broke. Property tests target the edge cases people tend to miss: write a statement you believe can never fail, then use a test library to generate thousands of inputs that try to prove otherwise. Round-trip encoding, 'an account balance can never go negative,' and idempotency all belong here. Differential tests suit large migrations: run the old and new systems on the same inputs, and any difference means a bug or a breaking change. Inputs should be randomly generated or sampled from real production traffic.

Rules that keep verification itself honest

Tests can be written, and verification can be gamed — so the team has laid down a few hard rules. Review rules live in REVIEW.md, read only from the main branch, so a change can never touch the standard that judges it. A failing check gets at most three rounds of fixes before a human is brought in. Every commit is stamped with how it was verified, including 'none,' so unverified work can be queried later. A new commit gate starts in observe mode, logging what would have been blocked, and only switches to enforcement once false positives look manageable. The team also keeps a 39-item reliability checklist, each item triggered by something readable from a diff, so only the few matching items are checked — no one has to memorize past outages.

This approach is not Anthropic-only. In internal testing, adding an independent grader that scored structured file generation against a rubric improved task success by up to 10 points. OpenAI recently invited four outside evaluation firms to inspect model alignment directly inside its labs — the same instinct: verification cannot rest on your own word alone. Ordinary teams can copy three things: write a verify skill for the repo so Claude actually runs the app before committing; stop after three failed rounds and bring in a human; and turn every bug that slipped through into a new rule, pruning stale rules on a regular basis. The verification layer itself needs to be maintained like any system — otherwise it soon becomes decoration.

Recommended Tools

More