One tester ran 1,872 prompts across Haiku, Sonnet, and Opus models to measure how often Claude Code loaded six public testing skills. With original descriptions, the skill fired in 15% of runs pooled. Pushy rewrites of the descriptions added only 5.2 percentage points, mostly from three of the six skills, missing the tester's preset threshold.
The most significant boost came from Superpowers, a Claude Code plugin. Adding the plugin's session-start hook — which instructs Claude to check its skills before any other task — raised the skill firing rate from 15% to 62% pooled across all models. The tester found that near-miss requests, which share a word with a skill but require different context, never fired their target in any of the three setups, totaling 0 of 972 runs.
A skill loads when the agent makes its first tool call. With many skills installed, Claude Code lists only names to conserve context, cutting descriptions short. The study concluded that description wording has limited impact, but the session-start hook dramatically improves skill utilization.
No near-miss request succeeded in triggering a skill across any model or description style.
Source: Towards Data Science · Summarized by HeadlinesBriefing