# How the Claude Code team verifies Claude's code

> Anthropic's Claude Code team, where Claude writes about 90% of the code, checks it with a verify skill that runs the app end to end. The team explained how in a post published 2026-11-16.

- Published: 2026-11-17
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/blog/claude-code-verification-tests/
- Answers: Claude Code verification tests
- Agent file (a short card for your agent): https://riglabs.ai/blog/claude-code-verification-tests/AGENT.md

---

You can give Claude Code a verify skill that starts and drives your app, so Claude checks its own work before you look at it: the same setup the Claude Code team at Anthropic uses. The team explained how it checks Claude's code on November 16, 2026, in [a post by Addy Osmani on claude.dev](https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/).

Claude writes and reviews about 90% of the team's code, and the team's main rule is to run a change end to end before trusting it.

## Why does verification matter so much with Claude Code?

Verification decides how much work you can hand off, per the post. Without a way for the agent to check its output, you can run one or two Claudes at a time, because you check everything yourself. With it, the team says, you can run dozens in parallel.

## Which tests does the Claude Code team recommend?

The post names three kinds of tests worth investing in, and says to pick end-to-end if you can only focus on one:

- End-to-end tests first. They follow a real user path (add to cart, check out, see the confirmation) and are hard to fool: the feature works or it does not. Unit tests can pass while the feature is broken.
- Property tests. You state something that must always hold ("a balance can never go negative", "decode after encode gives back the input") and a library such as Hypothesis or fast-check tries thousands of inputs to break it.
- Differential tests. For a migration, run the old and the new system on the same inputs and compare the results.

Tests must be fast and not flaky. A flaky test teaches the agent to rerun it instead of fixing the code.

## What does the team actually do?

The Claude Code team's own practice, per the post:

- A skill called verify, in a `.claude/skills/` folder next to the code, holds how to start and drive the app (the Chrome extension for web, an iOS or Android MCP for mobile, computer use for desktop). Claude uses it before committing.
- Claude runs the tests, build and linter before opening a pull request.
- If a check still fails after three rounds of fixes, Claude stops and brings in a person.
- A change can never edit the check that judges it: review rules are read from the main branch.
- A repeated failure becomes a hook or a CLAUDE.md rule, so it does not happen again.
- Their commit gate, which blocks a commit made without /verify or /code-review, has an observe mode that logs what it would block before it blocks anything.

## What to do today

Ask Claude Code, in a repository you work on:

```text
Create a skill called verify in .claude/skills/verify/SKILL.md. It should start
this app locally, drive the main user flow end to end, and report pass or fail
with what it saw. Then run it once. Fix the skill only if it fails to start or
drive the app. If the app itself fails, report it, and do not change the check
to make it pass.
```

Last checked 2026-11-17.

Sources:
- [claude.dev: How the Claude Code team thinks about verification, Addy Osmani, 2026-11-16](https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/)
- [@ClaudeDevs on X, 2026-11-16](https://x.com/ClaudeDevs/status/2122457608963395586)
