Claude Writes 90% of Claude Code: Its Verify Loop Explained

About 90% of the Claude Code team’s code is written and reviewed by Claude, Anthropic says, and on 16 November it published how the team checks that work, in an engineering post by Addy Osmani. A skill called verify, kept in the repo, boots and runs the app before Claude commits; if a check is still failing after three rounds, Claude stops and brings in a human; and a 39-item reliability checklist fires only the items whose trigger matches the change.

Our number: Claude checks only the items whose trigger matches the diff, “usually 3 to 10” of the 39, so each change runs 8% to 26% of the checklist, by our arithmetic. The volume is why: in Anthropic’s monorepo, changes landing per day went from about 2,500 in May to more than 30,000 in September, Anthropic says, 12 times, by our arithmetic, the growth behind the Scion source-control redesign we explained in October.

Business Pill · Verification Cost

An evergreen Business Engineer explainer on verification cost: output that is cheap to make is not always cheap to check. It teaches the general idea only and does not discuss Anthropic’s post.

The key insight: As we read Anthropic’s post, the scarce step is no longer writing the change but proving it works: Claude writes and reviews about 90% of the team’s code, Anthropic says, and the people are spent where checks fail, after three rounds, while only the 3 to 10 checklist items that match a diff run, 8% to 26% of the list, by our arithmetic.

How the Verify Loop Works

The post says running the app end to end matters more than tests. Claude is told to use the Claude Chrome extension for web changes, an iOS or Android MCP for mobile, computer use for desktop, or to run the service locally for backend work. The steps for starting the simulator or the app go in a skill called verify, kept in a .claude/skills/ folder in the repo, and “Claude knows to use the verify skill before committing changes”.

Anthropic says “This is how we test nearly all changes at Anthropic”. Failures decide where people look: “Nobody can read every change, so we use failures as one way to decide what needs a person”. Claude runs the checks and tries to fix what fails, and “If a check is still failing after three rounds, Claude stops and brings in a human”.

The people who own the code write its checks. Each item on the 39-item checklist has a trigger that can be read from a diff, a check, and the incidents or pull requests behind it; one example the post gives is “–resume loads old transcripts”. Anthropic says the design means “nobody has to memorize past outages”.

The Claude Code team's verification path as Anthropic describes it in its 16 November 2026 post, simplified; t
The Claude Code team’s verification path as Anthropic describes it in its 16 November 2026 post, simplified; the order of steps is our reading. Our arithmetic: 3 to 10 of the 39 checklist items is 8% to 26% per change.

Checks a Change Cannot Edit

The post’s rule is that “If a change can modify the check that judges it, the check is broken”. Code Review reads its REVIEW.md file from the base or default branch, never from the pull request’s own changes, and the rules that gate changes sit on a branch the change cannot modify. Reviews are reserved for commits, because “A passing review on an out-of-date commit means nothing”.

Cheap checks come first: Claude runs the tests, build and linter before opening a pull request. The team stamps each commit with how it was verified, including “none”, and runs a commit gate that can block a commit if the session did not contain /code-review or /verify, with off, observe and enforce modes. If commits have been blocked twice in a session, the next one is let through, “so Claude doesn’t get stuck”.

One Reviewer, and Misses Turned Into Rules

Anthropic says it does not add reviewers to improve review: a second, more expensive reviewer on top of the first “is often a worse approach than tuning the primary”. The tuning goes into a REVIEW.md that says what the reviewer should look for in that code, and later rounds build on earlier ones instead of re-checking everything.

Misses become rules in two ways. A pattern from human review comments and reverts must occur in at least two pull requests and be visible in a diff; three agents then try to disprove it, and a human approves it in a pull request. Separately, three bug finders are replayed against a pull request tied to a known incident; if at least two find the bug, it counts as detectable.

Flaky tests cannot be skipped, padded with sleeps or retried until they pass; if Claude cannot find the root cause, it deletes the test and writes down why.

The team also reviews the reviewers, tracking how many merged changes were reverted, how many changes a maintainer rejected before merging, and a size-matched subset of changes that stayed clean.

Bar chart: changes landing per day in Anthropic’s monorepo, about 2,500 in May 2026 and more than 30,000 in September 2026
Changes landing per day in Anthropic’s monorepo, about 2,500 in May and more than 30,000 in September 2026, as Anthropic’s 15 October code review post reports them; the monorepo is separate from the Claude Code repository. The 12 times multiple is our arithmetic.

What the Code Review Post Measured

The verification post builds on Anthropic’s 15 October code review post, which gives the outcome figures. From June to September, merged pull requests to Claude Code’s codebase more than doubled while the share that had to be reverted fell by more than 40 percent, it says. When Anthropic launched its Code Review research preview, the share of internal pull requests getting substantive review comments went from 16% to 54%.

The same post says one pull request went through more than fifty rounds of review and fixes overnight before its author stopped it, and that about 59% of what the reviewer kept flagging was code that had been there since its first pass. The fix was a round cap: after three rounds, the reviewer asks the author to step back and reconsider the approach. In separate internal testing on structured file generation, the new post says, a separate grader checking Claude’s work against a rubric improved task success by up to 10 points.

Anthropic is not the only company citing such a share; in our Proton piece, Proton’s CTO put AI’s share of the company’s October code at 90%.

The Structural Read

As we read the design, the checks are built so the code cannot grade itself: the review rules sit on a branch the change cannot modify, and a review counts only for the commit it read.

As we read the volume figures, the method answers a scale problem: Anthropic’s monorepo went from about 2,500 changes a day in May to more than 30,000 in September, 12 times, by our arithmetic, a pace at which, the code review post says, line-by-line human review is incredibly difficult if not impossible.

Addy Osmani, Anthropic engineering post, 16 November 2026

“Verification loops are the building block that lets you delegate more work to Claude.”

Three Implications

FOR ENGINEERING TEAMS Anthropic’s post says that without verification loops you can run only one or two Claudes at a time, and with them, dozens or hundreds in parallel.

FOR CODE REVIEW Anthropic says a second, more expensive reviewer on top of the first is often a worse approach than tuning the primary one with a REVIEW.md.

FOR QUALITY From June to September, merged pull requests to Claude Code more than doubled while the share reverted fell by more than 40 percent, Anthropic’s code review post says.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “Universal harnesses are likely to dominate highly digital, modular, and verifiable work.”

As we read Anthropic’s post, software is that kind of work, and the verification layer, the verify skill, the diff-triggered checklist and the checks a change cannot edit, is what makes it verifiable at volume; the framework predates the post and does not discuss it.

What Is Not Established

The roughly 90% share and the outcome figures are Anthropic’s own, about its own team; we have not seen the underlying data. The posts we read give relative changes for the Claude Code repositories (more than doubled, more than 40 percent) rather than counts of pull requests or a revert rate, and the 2,500 and 30,000 figures are for Anthropic’s monorepo, which the code review post says is separate from the Claude Code repository.

The up-to-10-point gain comes from internal testing on structured file generation, not from the Claude Code team’s code. The verification post describes practices; it does not report how often the three-round hand-off or the commit gate fires.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

The Business Engineer analysis of why the harness around the model, not the model itself, is where value pools in the AI stack.

Read the Map of AI →

The Bottom Line

Anthropic says Claude writes and reviews about 90% of the Claude Code team’s code, and its verification post sets out the checks around that work: a verify skill that runs the app, a hand-off to a human after three failing rounds, and a 39-item checklist of which usually 3 to 10 items fire, 8% to 26% per change, by our arithmetic. The post’s closing line puts the job of the people as “defining what “done” looks like and reviewing the evidence”.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The practices and figures come from Anthropic’s engineering post on how the Claude Code team thinks about verification, by Addy Osmani, published on 16 November 2026, which we read in full, and from Anthropic’s 15 October post on how the Claude Code team does code review, which we read in full as text. We found the story through a relay post on X by @alexmichael and checked every point against Anthropic’s posts.

The percentages of the checklist per change and the 12 times multiple are our arithmetic. We did not contact Anthropic for this piece.

Sources: Anthropic: How the Claude Code team thinks about verification, by Addy Osmani (16 Nov 2026) · Anthropic: How the Claude Code team does code review (15 Oct 2026)

Scroll to Top
FourWeekMBA