Auto-trigger vs slash-command skills: does auto-triggering actually fire?

status: pre-registered · protocol published 2026-07-07

Protocol pre-registered 2026-07-07. Results publish when all runs complete.

Hypothesis

Auto-triggered skills fire on under 80% of tasks they should handle when prompts are naturally phrased, and the miss rate is driven by description wording rather than task difficulty.

Method

Tasks20 prompts that should trigger a target skill (4 skills x 5 phrasings from direct to indirect), plus 10 adjacent prompts that should not.
ConditionsEach prompt in a fresh session; fire/no-fire recorded from the transcript. Slash-command control arm establishes the ceiling (invocation is deterministic).
RunsN=3 per prompt (90 trigger observations).
ScoringTrigger precision and recall per skill; phrasing-distance curve; false-fire rate on the should-not set.
PinnedSkill versions, exact prompt list, and session configuration pinned at registration.

Results

Empty until runs complete. This section is the only part of the page that will change, and the change will be logged in the changelog.

Limitations

Trigger behavior is model-version sensitive; results date quickly and will be marked with the tested version. Four skills is a sample, not the ecosystem.

Go deeper