Claude Code team says ~90% of its code is written and reviewed by Claude
Anthropic's Claude Code team disclosed in an engineering blog post (by Addy Osmani) that roughly 90% of its code is written and reviewed by Claude. Since no one can read every diff, the team relies on verification loops rather than human review as the safety net, enabling dozens to a hundred parallel Claude instances instead of one or two. minchoi and FinanceYF5 broke the methodology into 7 rules plus supporting engineering practices.
Confirmed
- The blog's core thesis: verification loops are the foundation of delegation—without them you run one or two Claudes in parallel; with them, dozens to a hundred.
- minchoi's 7 rules: (1) end-to-end tests first, simulating real user flows (add to cart, checkout, confirmation page); (2) place a verify skill in .claude/skills/ so Claude actually runs the change itself before committing; (3) property tests for invariants that must never break—e.g., balances never negative, migrations idempotent—using fast-check (JS/TS), Hypothesis (Python), or proptest (Rust); (4) for migrations, run old and new systems in parallel on identical inputs and treat differences as bugs or breaking changes, with noise normalization and an approved-differences list; (5) after the same check fails 3 rounds, stop and escalate to humans; (6) a change must never edit the checks that judge it—code review reads rules from the base branch; (7) turn recurring AI mistakes into verifiable rules derived from human review comments, PR rollbacks, and past incidents, with new rules validated by an agent plus human approval.
- FinanceYF5's additional details: the team maintains 39 reliability checks, each with trigger conditions tied to historical failures; code review uses REVIEW.md to tell the reviewing agent what to focus on rather than stacking more review models, with off/observe/enforce gating; engineers shift from line-by-line review to defining "done," with agents supplying test results, screenshots, and verification evidence.
Why it matters
- A reference template for scaling AI coding teams: when AI output far exceeds human-reviewable volume, rules like "failures decide human intervention points" and "changes can't tamper with their own judges" are more sustainable than adding more review models.
- The concrete practices (verify skills, property-testing tool choices, differential migrations, three-level gating) are directly adoptable by engineering teams using Claude Code or other agents.
2026-11-18 ~ 2026-11-19 · 15 related posts
Primary sources
- Three tests the Claude Code team swears by: E2E, property, differential — FinanceYF5 · 2026-11-18
- [source] Claude Code team: stop after 3 failed rounds, hand off to humans — FinanceYF5 · 2026-11-18
- Claude Code review design: REVIEW.md, base-branch rules, 3-mode gating — FinanceYF5 · 2026-11-18
- How the Claude Code team turns recurring AI mistakes into validated checks — FinanceYF5 · 2026-11-18
- Claude Code team's playbook: engineers define "done" while agents provide verification evidence — FinanceYF5 · 2026-11-18
- [source] Claude Code team says Claude writes and reviews ~90% of their code, aided by 7 verification rules — minchoi · 2026-11-19
- Claude Code team runs dozens of Claudes in parallel: 7 verification rules, starting with E2E tests — minchoi · 2026-11-19
- Give Claude a way to run the change, not just the tests: verify skill in .claude/skills/ — minchoi · 2026-11-19
- Property tests for invariants: throw thousands of generated inputs at rules that must never break — minchoi · 2026-11-19
- Claude Code rules: property tests for invariants and old-system-as-oracle migrations — minchoi · 2026-11-19
- Claude Code team says AI writes ~90% of code: after 3 failed fix rounds, hand off to a human — minchoi · 2026-11-19
- Claude Code rules: 3 failed fixes then a human, and checks a change can't edit — minchoi · 2026-11-19
- Claude Code verification rules: differential tests, REVIEW.md isolation, and mined rules — minchoi · 2026-11-19
- [source] Claude Code team blog: verification loops let coding agents prove their work — minchoi · 2026-11-19
1 near-duplicate retellings: FinanceYF5