Don't let Claude grade its own homework

Short summary

Anthropic's Claude Code team uses two separate AI agents for code review—one writes, another reviews skeptically—to prevent the trap of AI grading its own work. By refreshing context between writing and review phases, issues surface earlier in development. The practice layers AI review during development, at PR level via CI, and human judgment for final approval.

• •Use separate AI agents for writing and reviewing code to avoid self-validation bias
• •Kill your session between writing and review to get a fresh context window
• •Layer AI review (dev-time), PR-time (CI), and human judgment for comprehensive coverage

Generated with AI, which can make mistakes.