skill-comply

You can verify agent definitions and rules are followed by auto-generating behavior specs from MD files, running agent traces with varying prompt strictness, and classifying tool calls against those specs using an LLM.

The best way to measure compliance is by executing end-to-end Claude Code workflows where scenarios are run, tool call traces are captured, and temporal ordering is checked deterministically to produce a self-contained compliance report.

You auto-generate behavior specs from markdown files by parsing the MD data to create spec sequences, which are then used to generate scenarios with supportive, neutral, and competing prompt strictness levels for testing.

Yes, you can classify tool calls against specs without regex by using an LLM to evaluate the captured tool call traces, ensuring behavioral compliance is measured through semantic understanding rather than rigid pattern matching.

Yes, it works with existing skills and agent setups by running a full compliance check against your current definitions to quantify whether they execute as intended and to identify any behavioral drift caused by recent changes.

A limitation of checking temporal ordering deterministically is that it requires accurately captured sequential tool call traces from the agent runs; any asynchronous or parallel execution outside the captured trace may not be validated.