Claude writes about 90% of Claude Code. The spicier part? it also has to prove the code works before a human bothers looking.
That's the gist of a new post on the claude.dev blog, "How the Claude Code team thinks about verification", which hit 283 points and 227 comments on Hacker News. The team says the vast majority of its code (~90%) is written and reviewed by Claude, and at that scale the thing that matters is how the agent checks its own work with evidence.
My hot take: the post is mostly classic testing discipline wearing a new hoodie. And that's exactly why it's worth reading.
ok so what's the actual recipe
The core argument is simple. Verification loops are what let you delegate. Without them you can run one or two Claudes at once because you're checking everything yourself, with them the post says you can run dozens or hundreds in parallel.
They back it with one internal number: on structured file generation, a separate grader checking Claude's work against a rubric improved task success by up to 10 points over a standard prompting loop, with the biggest gains on the hardest problems.
Then the three test types they rate most:
- End-to-end tests. If you only pick one, pick this. Either the whole feature works or it doesn't, which makes them hard to fool. Unit tests can pass while the feature is still broken.
- Property tests. Write a statement you think can't fail, then let fast-check, Hypothesis or proptest throw thousands of inputs at it. The point is edge cases that aren't made with the same bias as the agent that wrote the code.
- Differential tests. For migrations: run old and new systems on the same inputs and diff the results.
The practical bit is a skill called verify that lives in .claude/skills/, close to the code it covers. It tells Claude how to boot and drive the app, and Claude knows to use it before committing. The team says this is how they test nearly all changes at Anthropic.
the boring bits are the point
ngl the most useful lines in the post aren't about test types at all. They're about keeping the checks honest.
"If a change can modify the check that judges it, the check is broken." So Code Review reads REVIEW.md from the base or default branch, never from the pull request. That's lowkey the whole game. An agent that can edit its own grader will eventually grade itself an A.
Other rules I'd steal tomorrow:
- if a check still fails after three rounds, Claude stops and brings in a human
- flaky tests can't be skipped, padded with sleeps or retried until green. If Claude can't find the root cause it deletes the test and writes down why
- every commit gets stamped with how it was verified, including "none", so unverified commits are queryable
- a commit gate with off, observe and enforce modes, and after two blocks in a session the next commit goes through so Claude doesn't get stuck
There's also a 39-item reliability checklist where each item has a trigger readable from a diff. Claude only checks the items whose trigger matches, usually 3 to 10. imo that's the smartest idea in the whole post, it turns past outages into something nobody has to memorize.
HN said "just use Playwright" and they're half right
The first comment in the Hacker News thread called it the work of a team with unlimited tokens, arguing Playwright gives the same quality with deterministic results at near zero marginal cost.
Fair, but kinda misreading it. The post never says skip deterministic tools. It literally says to make the verify skill your smoke test harness and make it fast and reliable, and to run cheap checks (tests, build, linter) before opening a PR. A couple of commenters pointed this out too: nothing stops your verify skill from calling Playwright. Tbh that's probably the move.
The cost fight got fun though. One commenter guessed $500 of tokens per release, 2 to 3 releases a week, and called it a drop in the ocean. Another did the math at 2,000 to 3,000 PRs a week and landed on $6M a month. Those are commenter estimates, not anything from Anthropic, so treat them like vibes with a calculator.
The sharper criticism imo was about churn. When an oracle test fails, one commenter said, the model can rewrite the test, copy buggy behavior to match it, or build a quirks mode. That's real, and it's exactly why the "checks live where the change can't touch them" rule exists.
stealing this for your own agents
Even if you're not merging hundreds of PRs, the pattern carries over to any agent that does real work.
Say your coding agent opens PRs on GitHub and runs its own checks. Same principle as REVIEW.md: it shouldn't have the power to approve or change whatever judges it. If you route it through MewCP you pick which tools that agent gets, down to actions, and anything you switch off is stopped before it reaches the app. so the agent that writes code doesn't also hold the keys to wave it through.
And setup isn't the hard part here. Adding MewCP to Claude Code is one command, the hard part is deciding what "done" looks like.
stuff the post doesn't tell you
- The "up to 10 points" figure is from internal testing on structured file generation. No benchmark details, no public eval.
- No cost numbers. The token bill HN is fighting about is pure guesswork from commenters.
- We don't know how often the escape hatches fire, like the third-commit pass after two blocks, or how many tests get deleted as flaky.
The ending lands well though. Your job shifts from reading every diff to defining what "done" means and reviewing the evidence. Not me realizing that's just being a good tech lead, with more tabs open.
Sources