Anthropic is testing whether Claude Code can handle the daily maintenance of the company's own software. According to Boris Cherny, the Anthropic engineer who created Claude Code, the results are "surprisingly positive."

The setup

As Cherny describes it, Claude runs daily routines via Tag through a dedicated Slack channel called "proj-claude-maintains-apps." The scope covers all of the company's platforms: iOS, Android, desktop, web, CLI and the Agent SDK.

The point is to stop tying up developers with repetitive maintenance work — the unavoidable side of software that nobody volunteers for: tracking down the root cause of a crashing screen, stripping out code that no longer runs, merging two abstractions that do nearly the same thing.

Twelve routines

The specialised maintenance routines Claude runs include:

  • Crash Fuzzer: opens apps in a simulator, taps around randomly to trigger crashes, analyses the root cause and creates a fix
  • Dead-Code Remover: strips statically unreachable code; for suspicious code it first adds logging, then checks the next day whether the code is genuinely unused
  • Dup Unifier: scans the codebase for similar but slightly different abstractions and proposes merging them
  • Logic Simplifier and Logic Bug Fixer: simplifies nested business logic and models complex logic to find errors
  • Useless Test Pruner: removes tests that can never fail
  • Flaky-Test Fixer: analyses and repairs unstable CI tests
  • Abstraction Improver and Abstraction Police: simplifies overengineered abstractions and fixes layer violations in the architecture

The dead-code remover's method is notable: rather than deleting code it is unsure about, it first leaves a log line there and confirms the next day that the code really never runs. That is the same cautious approach a careful human engineer would take.

No prompt engineering

Cherny shared some of his prompts in Slack, and there is no elaborate prompt engineering going on. He tells Claude in plain language to start daily crash-fuzzing routines on iOS, Android and desktop, to use real apps rather than mocks, to trigger crashes, and to open pull requests with fixes.

388 requests, 180 merges

Claude created 388 pull requests across Anthropic's repositories in the first few weeks. After a combination of automated Claude Code review and human review, 180 were merged — a rate of roughly 46 percent.

Claude usually gets the pull requests right on the first try, Cherny says. When it does not, the team tweaks the routine so the AI does better the next day; that tuning sometimes takes a few days. Anthropic is now looking at ways to speed up the merge process.