Anthropic's code-review skill helps Opus 5.5 fix 5.5 more planted bugs at lower cost
PawelHuryn · x · 2026-10-17
Pawel Huryn's Bug Hunt Bench benchmark: 105 planted bugs across two real codebases, blind-graded with the same model, effort, and judge. Claude Opus 5.5 at high effort with Claude Code's built-in /code-review skill fixed 37.5/105 planted bugs on average vs 32.0 without (+5.5), with both skilled runs beating all three unskilled runs.
- Cost: $15.00 with the skill vs $22.25 without; fewer, smaller model calls on the larger codebase
- Bonus: real unplanted bugs fixed doubled — 24 vs 10.7 on average (never counted toward score)
- Mechanics: at high level the skill runs eight review angles (line-by-line scan, removed behavior, callers/callees, reuse, simplification, efficiency, root cause, CLAUDE.md rules) in one agent, up to six candidates each, then dedups and fixes; no subagents or verification at this level (those run only at max)
- Caveats: two runs with skill vs three without; repeat runs of one setup differ by 5/105 at median
Huryn also notes his own improved /code-review skill, which focuses model attention on bug classes models commonly miss, and plans a follow-up comparison against Anthropic's version with subagents enabled.
More from coding & agent
- Open-source coding agent OpenCode hits 214k GitHub stars with free DeepSeek V4.1 Flash built in — alex_verem · 2026-10-17
- AICre8 MCP Server lets agents build and deploy web projects in a sandbox — modelcontextprotocol · 2026-10-17
- Energetica MCP connector serves 40+ curated Argentine oil & gas tables — modelcontextprotocol · 2026-10-17
- Community ports DeepSelect to Hopper with CuTe DSL in TokenSpeed PR — zhyncs42 · 2026-10-17
- ComfyUI node turns illustrations into manga panels using Qwen-Image 2.1 Turbo — KnowledgeBig4364 · 2026-10-17
- Open-source CLI scans MCP configs for poisoned tool descriptions before agents run them — StanislavMZ · 2026-10-17