What problem does it solve? Teams write skills, rules, and agent definitions for Claude Code but have no way to verify whether agents actually follow them in practice. This Skill quantifies real behavioral compliance instead of assuming instructions are obeyed. ## Core Features & Use Cases - Automatic Spec Generation: Extracts an expected behavioral sequence (ordered steps with required flags) from any SKILL.md, rule, or agent definition file using an LLM. - Three-Level Scenario Testing: Generates supportive, neutral, and competing prompts for the same task to measure prompt independence — whether a rule is followed even when the prompt does not ask for it. - LLM-Based Trace Classification: Runs scenarios via claude -p with stream-json output, classifies each tool call against spec steps by meaning (not regex), and checks temporal ordering deterministically. - Use Case: After adding a new TDD rule to your project, run a dry run to inspect the generated spec, then execute full scenarios to get a compliance report showing which steps agents skip and whether hook promotion is recommended. ## Quick Start Run the skill-comply analysis on my testing.md rule file to measure whether agents actually follow it.
