Anthropic Details How Claude Writes 90% of Claude Code and Automates Code Review

Anthropic engineers reveal that Claude writes approximately 90% of Claude Code's own code, necessitating a fundamental shift in how code changes are reviewed and merged. By moving from human-centric review gates to continuous, machine-assisted verification, the team achieved a 2x increase in merged PRs and a 40%+ reduction in revert rates between June and September. Daily changes in the Anthropic monorepo surged from ~2,500 in May to over 30,000 in September.

The 7-Stage Review Lifecycle

The new process redefines code review not as a final gate, but as continuous feedback similar to spellchecking. The lifecycle includes:

1. Pre-PR Verification: Claude runs tests, builds, and linters. A fresh-context subagent reviews the diff. Internal skills check for bugs, simplify diffs, and ensure UI guideline compliance.
2. AI Review for All PRs: An AI agent analyzes the diff against the whole repo and PR intent. This increased substantive review comments on internal PRs from 16% to 54% after the research preview launch. Findings are tagged as Important, Nit, or Pre-existing.
3. Security Review: A blocking check using open-source security review tools and /security-review commands. Open findings prevent merging.
4. Automated Checks: CI runs linting, type checks, and tests across all platforms. PR templates require testing and undo instructions.
5. Risk-Based Human Approval: A classifier determines the risk tier:
   • Low (Docs, flags): No human approval needed if automated checks pass.
   • Medium (Product changes): 1 human approver.
   • High (Critical code): 2 human approvers.
   • Constraint: Agent approvals do not count as human approvals; agents cannot approve their own work.
6. Merge Queue & Staged Release: Changes go through a merge queue, are used internally by Anthropic staff, and released to latest before stable.
7. Detection & Rollback: Feature flags allow quick disabling of features. Weekly on-call and incident reviews analyze which layer failed to catch issues.

Convergence and Efficiency Gains

A major challenge was convergence loops, where an agent author and reviewer would cycle indefinitely (e.g., one CLI PR had 50+ rounds). Analysis showed 59% of repeated reviewer comments were about pre-existing code, not new bugs.

To fix this, Anthropic implemented round caps (3 rounds max), diff-only re-reviews, and duplicate finding bans. Results for internal PRs:

• Passing in 1 round: Increased from 65% to 78%.
• 5+ rounds required: Dropped from 8.9% to 1.5%.
• Average review rounds: Decreased from 2.08 to 1.32.

Human-in-the-Loop Responsibilities

Humans are no longer reading every diff. Instead, they focus on five areas where machines lack judgment:

1. Strategic Direction: Deciding if a feature should be built (scope, taste).
2. Risky Code: Billing, auto-updaters, and CI workflows.
3. Architecture: Setting patterns for others to follow.
4. Reading Evidence: Interpreting test outputs, screenshots, and complex reviewer findings.
5. Ownership: Humans remain responsible for the outcome of all changes, regardless of who (or what) wrote the code.

Adoption Guidelines

Anthropic suggests other teams adopt this by:

• Adding CLAUDE.md and REVIEW.md to define project conventions and review rules.
• Using /code-review and /security-review locally.
• Implementing risk-based approval tiers and strict branch protection for critical paths.
• Using incident data to build eval sets for continuously improving AI reviewer performance.