AGENTS.md outperforms skills in our agent evals

Takeaways

Vercel evals on Next.js 16 APIs: an ~8KB compressed docs index in AGENTS.md hit 100% pass rate; skills capped at 79% even with explicit “use the skill” instructions.
Skills were not invoked in 56% of cases—same as no-docs baseline (53%)—agents often ignore available tools.
Explicit trigger wording in AGENTS.md raised skill usage to 95%+ and pass rate to 79%, but phrasing was fragile; small wording changes shifted behavior dramatically.
Persistent root context (AGENTS.md / CLAUDE.md) beats on-demand skill loading when framework APIs are stale vs. model training data.
Version-matched doc indexes solve both “model suggests too-new APIs” and “model falls back to old patterns.”

Validates this vault’s AGENTS.md + rules architecture over stuffing everything into skills—keep durable, always-on context tight and version-aligned.

What is the right size/compression tradeoff for vault AGENTS.md as skills multiply?