Claude Skills Fire Rate Jumps to 62% With Session Hook

A tester named in the piece ran 18 prompts across Claude's Haiku, Sonnet, and Opus models to answer a question Anthropic's own documentation hedges on: how often do installed Skills actually fire. The setup was simple. A workspace held a public skill called python-testing-patterns, built to trigger on requests to write pytest tests with fixtures.

The test prompt asked for exactly that, sharing five words with the skill's own description. By any reasonable matching logic, it should have loaded every time.

It didn't. Across all three models and both the original and a "pushy" rewritten description, the skill failed to fire on a near-miss prompt in run after run. Instead, each model improvised its own workaround: writing test files from scratch, saving functions to new locations, checking for pytest manually.

Claude Code's interface gave no error, no warning, nothing to flag the miss. The only way to catch it was combing through the transcript line by line for a Skill call that never came.

That gap between what a skill description promises and what actually gets invoked is the subject of this testing exercise, and the numbers that came out of it are worth sitting with before explaining how they were measured.