dogfood-skill-testing
Dogfood Skill Testing
Overview
A bad SKILL.md doesn't throw an error — it just never gets invoked. A skill defect is rarely a crash; the dominant failure mode is silent. The skill never fires, fires for the wrong request, or produces valid-looking output that is domain-wrong — and a green grader or a passing structural check sees PASS while the skill is broken in practice.
This skill fills the black-box exploratory behavioral gap that its
siblings leave open. skill-judge scores a SKILL.md statically (reads
the file, never runs it). skill-creator-advance runs a white-box
conformance eval against the author's own known prompts. distill-sessions
mines past session logs for activation telemetry. None of them unleashes
a fresh agent that does not know the author's intent and actively
probes a skill-in-development to surface the unanticipated defects no
author-written test prompt covers.