You wrote a skill (Tip #8), placed it right (Tip #9), but the agent doesn’t use it. You’re not alone.
Vercel’s engineering team ran rigorous evals on this. With default settings, the agent never invoked the skill in 56% of cases — even though it was available. Adding explicit instructions (“explore the project first, then invoke the skill”) raised usage to 95% and pass rate from 53% to 79%. But an 8KB docs index embedded directly in AGENTS.md hit 100% — no skill invocation needed.
Their theory: passive context (AGENTS.md) has no decision point — the info is just there. Skills require the agent to decide “should I look this up?” and current models aren’t reliable at that step.
The practical takeaway: use skills for tasks you invoke explicitly — deploy, release, code review checklists. For general knowledge the agent should always have (coding patterns, API docs, framework conventions), put it in AGENTS.md instead. Horizontal knowledge → instruction file. Vertical procedures → skills.
đź’ˇ Try this: In Copilot CLI, run /skills list and check descriptions. Are they specific enough for the agent to know when to trigger? Vague descriptions = missed activations.
Experienced skills not triggering? Found a workaround?