Boris Cherny shared an interesting experiment on his X account.
He leads Claude Code at Anthropic. What he tried over the past few weeks was handing the app maintenance people used to do over to Claude entirely.
A Maintenance Team Run From One Slack Channel
The setup was simple. He created a Slack channel called proj-claude-maintains-apps, and inside it had Claude repeat a few routines every day across iOS, Android, desktop, web, the CLI, and the Agent SDK.
- Crash fuzzer: operates the app in a simulator to provoke crashes, then finds the cause and fixes it
- Duplicate consolidator: finds code abstractions that are similar but slightly different and opens consolidation PRs
- Dead code remover: deletes unreachable code and plants logging in suspect code to check the next day whether it really is dead
- Abstraction cleaner: fixes leaky abstractions
Over three weeks, these routines opened 388 PRs. Of those, 180 passed both Claude's own code review and human review and were merged (original post, published at 6:27 a.m. KST on August 14, 2026).
Cherny said the main model used was Opus, and that other models were added only on a few tricky routines.
A 46% Merge Rate: How Is the Noise Filtered?
180 of 388 is under half merged. A question in the comments targeted exactly this point.
Cherny's answer was plain. "It's just that I haven't reviewed them yet, and the noise rate is roughly 1 in 50. If something turns out to be noise, I retune that routine."
In other words, most PRs not merged are a matter of review backlog rather than code quality, and the rate of real misfires (noise) is low. On the human review burden, he said, "The PRs are small and self-contained, so they're quick to review."
Start With Small Maintenance as Routines
What makes this case interesting is not "write one great prompt and build something big" but that small, repeatable maintenance tasks were turned into routines. Three takeaways for dev teams:
First, break work into units that are as small and self-contained as possible.
Second, treat a failing routine as something to tune, not discard.
Third, keep human review, but manage PR size so it doesn't become a bottleneck.
Cherny told people who want this workflow to build routines themselves in Claude Code or at claude.ai/code/routines.
The 388 Figure Is a Personal Account
This experiment comes not from an official Anthropic announcement or blog post but from a first-person account in the X post of the person in charge. The figure "180 merged out of 388" is a value Cherny reported himself and has not been re-verified through an outside audit or official statistics.
These routines are also configured for Anthropic's internal tools and repository environment, so they can't be moved to another organization as is without setup. If you're considering adopting this, it's safer to measure your noise rate first with a small pilot.