2026-11-16 · America/Los_Angeles · 社区动态 · #17
How we measured MCP tool discovery separately from skill use and answer quality
I maintain ai-rulez, a free MIT-licensed tool for agent configuration. I’m sharing a reproducible Claude Code experiment because a single success score hid several different failure modes. The fixture has eight skills. A conversation moves between two sibling directories with identically named worksheets but different workflows. For each positive turn we record three things independently: whether find_skill received the right path, whether load_skill actually returned the right body, and whether the response contained the expected code. The negative controls ask an unrelated question or ask…
热度 37.2 / 100;排名与评分保留该期记录。