How to Review AI-Written Code: Lessons From the Claude Code Team and Hacker News

Описание и применение

Проверка кода, генерируемого ИИ, становится эффективной, если сначала контролировать дизайн, а затем автоматизировать проверку. Команда Claude Code внедрила семиступенчатый процесс, включающий автоматический AI‑ревью, безопасность, линтеры и риск‑ориентированное человеческое одобрение, что сократило количество раундов ревью и повысило стабильность кода.

Задача

Reviewing AI‑written code is hard because line‑by‑line human review does not scale and automated tests can miss shared blind spots.

Результат

The pipeline reduced review iterations, increased PR pass rate, lowered revert rates, and doubled merged PRs while maintaining code stability.

Рабочий процесс

• 1. Agent runs tests, build, and linter locally before PR creation.
• 2. Fresh‑context subagent reviews the diff against the repo and PR intent.
• 3. Automated security review runs; PR cannot merge if any finding remains.
• 4. Blocking automated checks (lint, type, tests, build) run on every PR.
• 5. Human approval scaled by risk tier (low, medium, high).
• 6. PR enters merge queue with feature flags and gradual rollout.
• 7. Feature flags enable fast rollback and weekly on‑call monitoring.

Доказательства

The share of pull requests passing after just one review and fix round increased from 65% to 78%. Pull requests requiring five or more rounds dropped from 8.9% to 1.5%.

Источник

The first layer verifies the work before anyone else looks at it. Claude runs tests, the build, and the linter and shows the output.

Источник

The same Tokely API key also accesses OpenAI and Google models, such as GPT-6.1 Sol and Gemini 4 Argon, through the OpenAI-compatible Chat Completions endpoint.

Источник