I Can't Tell If the Model Matters

Short summary

A six-month experiment testing Claude Code and Gemini for code review reveals that repository context and prompt quality dramatically outweigh model lineage in code review effectiveness. First-party tooling with full repo access and adversarial prompts caught critical security issues that diff-only approaches missed. Model choice appears secondary to harness quality.

• •Context and prompt quality matter far more than model lineage for code review
• •First-party tooling with repo access catches security issues that isolated diffs miss
• •Harness design (setup, access, prompts) dominates model selection

Generated with AI, which can make mistakes.