When you install skills for your AI coding agent, you'd expect the agent to actually use them. But a lot of the time, it doesn't. Unless you explicitly tell it to. That friction got annoying enough that I built something about it.

Turns out Vercel ran evals on this exact problem. They tested skills (on-demand retrieval) vs embedding a docs index directly in AGENTS.md so it's always in context. Default skills performed the same as having no docs at all, 53% pass rate either way. Passive context in AGENTS.md hit 100%.

The reason is obvious in hindsight: with passive context, there's no decision point. The agent doesn't have to choose to load anything. It's just there.

Full writeup: AGENTS.md outperforms skills in our agent evals

What Skilldex Does

Skilldex takes your installed skills (Claude Code, Cursor, Windsurf, Copilot, etc.), reads their metadata, and writes a compact index into your AGENTS.md or CLAUDE.md ; whatever passive context file you use.

npx skilldex init

Scans for skills, you pick which ones to include, and it inserts a managed block into your target file. The agent sees everything on every request without you having to prompt it.

Other commands: add, remove, sync, list. Multi-target if you use multiple agents:

skilldex init -t AGENTS.md CLAUDE.md

Extensions

There's also an extension for both VS Code and Cursor. It adds a sidebar tree view showing all your indexed and available skills, with context menu actions for add/remove/sync. No terminal needed. Grab it on the VS Code Marketplace (works in Cursor too).

Source at github.com/nafisazizir/skilldex.