A Session-Start Hook Raised Claude Code Skill Firing From 15% to 62%

A large-scale test evaluated how effectively Claude's AI models—Haiku, Sonnet, and Opus—trigger specialized code skills in response to different user prompts. Making skill descriptions more "pushy," as officially recommended, provided only a minor boost, increasing the trigger rate for rephrased requests from a baseline of 15% to just 20%. A far more effective strategy involved using a session-start hook, which instructs the AI to review its available skills first, skyrocketing the success rate to an impressive 62%. Throughout the extensive testing, the models proved highly precise, never once incorrectly activating a skill when given a "near-miss" request designed to be similar but require a different tool.