AI Aulis

AI NEWS IN ENGLISH · 17 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 17 November 2026Paper

Claude Code team: Claude writes better code when it can verify its work

AI-generated contentWritten by AI

claude.dev, x.com · 2 sources · Sources published: 16–17 November 2026

Anthropic’s Claude Code team says a coding agent can take on more work when it can test its own changes. The team describes the tests it prefers and how it stops changes from modifying the checks.

Anthropic’s Claude Code team explained on 16 November, on the claude.dev blog and its X account, how Claude checks its own work in the Claude Code coding agent. Addy Osmani, who wrote the post, said about 90 per cent of the team’s code is written and reviewed by Claude.

The team’s reasoning is practical. If the agent cannot run checks itself and fix its work based on the results, a human has to check everything. Then only one or two Claudes can run in parallel. When the agent checks and fixes its own work, dozens or hundreds can run, the team said.

In the company’s internal testing, the task was to generate structured files. A separate grader that checked Claude’s work against a rubric improved task success by up to 10 points over a standard prompting loop. The biggest gains came on the hardest tasks. The figure was measured on this task type.

Three types of test

Osmani recommends end-to-end tests first. They follow a user’s real path, for example from shopping basket to order confirmation. Unit tests are faster but can pass even when a feature is broken.

A property test states a rule that must always hold, such as ‘the balance cannot go negative’. A testing library such as Hypothesis tries to disprove the rule with thousands of inputs. The inputs do not carry the same bias as the code the agent wrote.

A differential test runs the old and new systems on the same inputs and examines the differences in their results. It suits large migrations where behaviour must stay unchanged.

Flaky tests teach the agent to rerun a test instead of fixing the code, Osmani said.

Checks out of reach of changes

The team stores instructions for starting and running the app in the repository as a skill called verify, which Claude uses before a commit. Claude tries out a change as a person would: a web app through a Chrome extension, a mobile app through an iOS or Android MCP, and a backend service locally.

If a check still fails after three rounds of fixes, Claude stops and brings in a human. A change must not modify the check that evaluates it. Code Review reads its review instructions, REVIEW.md, from the base branch, not from the pull request’s changes.

The team’s commit gate can block a commit if /verify or /code-review has not been run in the session. A new gate starts in an observation mode, which logs the commits it would block but lets them through. Once the gate has blocked two commits in the same session, it lets the next one through so that Claude does not get stuck. The team acknowledges that Claude Code is not perfect.

Claude must not skip a flaky test, add waits to it or rerun it until it passes. If Claude cannot find the root cause, it deletes the test and records the reason.

Previous coverage

Report an error in this story