Skill-trigger reliability across 25 paraphrased requests

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

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

Hypothesis

For a single well-regarded skill, fire rate degrades predictably as request phrasing moves from the description's vocabulary toward natural task language, and a description rewrite following the trigger-test protocol recovers most of the loss.

Method

Tasks1 skill x 25 paraphrases graded by phrasing distance (vocabulary overlap with the skill description), from verbatim to fully indirect.
ConditionsPhase 1: original description. Phase 2: description rewritten per the Agentiquette trigger-test protocol, same 25 paraphrases.
RunsN=3 per paraphrase per phase (150 observations).
ScoringFire-rate curve by phrasing distance, before and after; the recovered-loss fraction is the headline number.
PinnedSkill version, both descriptions, the paraphrase set with distance scores, 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

One skill; the curve's shape may be skill-specific. Phrasing distance is a constructed metric, published with the data so it can be criticized.

Go deeper