rule-quality-evaluator

Evaluating agent instructions involves static extraction and seven-property scoring to identify noise, missing intent, and weak scope in your rules. This process audits instruction files to pinpoint why steering fails and highlights top improvements for behavioral impact.

Auditing .cursorrules or CLAUDE.md files requires structural diagnostics to detect redundancy with linters and map coverage across architecture, testing, and security. This identifies unfalsifiable guidance and scope mismatches that reduce instruction quality.

Testing behavioral impact is possible by generating targeted coding tasks and assertions. An evaluation harness measures whether the agent instructions actually change model behavior, rather than just evaluating how the rules read on paper.

Checking agent rule coverage involves structural diagnostics that map instructions across architecture, domain, conventions, integrations, operations, testing, security, performance, and tech debt. This exposes gaps where behavioral guidance is missing.

Agent instruction files fail to steer behavior due to vague rules, redundancy with existing CI types, or scope mismatches. Static rule scoring identifies these unfalsifiable guidelines and missing intents to improve overall effectiveness.