Code Audit

Multi-lens audit system that uses parallel sub-agents with directed checklists to assess the structural health of a codebase. Built on research showing that LLMs find significantly more issues when given specific, scoped checklists than when asked to "review thoroughly" (BitsAI-CR: 75% precision with structured rules; ECSA: 64%→82% precision with prompt detail; SWRBench: +43.67% F1 with multi-review aggregation).

Cost note: Token and runtime cost scales with scope and the number of active lenses. For large repositories, narrow the audit scope before running.

Model note: This skill performs deep analytical work. Use the most capable model available.

Core Principles

1. Context before judgment — Map the codebase before auditing anything. The scout builds the map, the lenses investigate.
2. Directed attention — Each lens has a specific checklist. Generic scanning produces surface-level results.
3. Convention-aware — Audit against this repo's actual patterns, not generic best practices.
4. Evidence-based — Every finding must include file location, code evidence, and impact. No vague opinions.
5. High signal — 5 actionable findings beat 20 vague ones. Prefer precision over volume.
6. Pull, don't push — Sub-agents read files themselves from disk. Never paste entire file contents into prompts.

Workflow Routing