writing-skills

To prevent agents from bypassing discipline-enforcing process rules, identify exact rationalizations they use under time, authority, sunk-cost, or exhaustion pressure, then encode explicit negations and red flags directly into the skill documentation to close those loopholes.

Writing YAML discovery metadata requires high-signal name and description fields where the description specifies triggering symptoms rather than summarizing workflow, ensuring better skill selection and discovery by other agents.

Agents bypass discipline-enforcing process rules because unvalidated skill instructions contain loopholes that allow rationalization under pressure, meaning the rules fail when agents face time, authority, sunk-cost, or exhaustion pressures during task execution.

Yes, you can use a TDD approach to harden documentation-based agent skills by running a baseline scenario to watch the skill fail first, then updating the skill content until that same scenario passes compliance verification.

The best way to enforce agent compliance is using third-person compliance language within skill instructions combined with a RED-GREEN-REFACTOR evaluation loop that pressure-tests the rules and re-verifies that explicit loophole counters prevent bypass attempts.