Hey PH 👋 I'm Mohamed. The inspiration was stupidly simple: I kept building agent skills that were technically fine — and never got used. Not because the logic was broken, but because the agent never opened them. Runtimes like Claude Code decide which skills to load based on name + description alone, and almost nobody tests that gate. So I built the tester I wished existed. The process evolved more than I expected. v0.1 was just a parser: is the description well-formed? Useful, but it couldn't answer "will this actually fire?". So I added the behavioral layer — a labeled suite of should-trigger / should-NOT-trigger utterances judged against name + description only, turning verdicts into a diagnosis, a rewritten description you can paste, and a regression suite you re-run after every model update. One thing I refuse to fake: results are a simulation of triggering judgment, not a runtime guarantee — that honesty rule is written into the docs themselves. What's next (open issues on the repo): an independent evaluator, repeated trials with confidence intervals, and collision analysis between rival skills. I'd love feedback — especially from people maintaining skill collections: what would make you trust a trigger report?

trigger-doctor

Find out why your AI agent skills never fire

Find out why your AI agent skills never fire