A 60-second arcade trailer of a real skill-routing run: 429 agent skills down to 2

I build jeo-skills, an open-source catalog of agent skills (429 in 10 categories) with a small router CLI. The hard part of a library this big is context: just listing every skill's name and description is roughly 53K tokens (my estimate: characters / 4).

So I turned one real routing run into a 60-second arcade trailer. First, local keyword search (BM25, no LLM) shortlists candidates; for this task the best skill (video-production) wasn't in its top 10. Then Jev (TypeSafe's decision model, optional, hosted API) judges by meaning: three gate questions average 0.72 against a 0.30 gate, the catalog is split into two shards, and 6 finalists come back.

A 0.30 fit cutoff keeps 2 skills (video-production 0.60, openmontage 0.37). Estimated listing context: about 53,126 tokens down to about 448. Each sub-task is then routed separately: sound to rfxgen, render to video-production, QA to audit-verify-explain-grade-5. For sprites Jev said "no confident match", so those were drawn by hand.

What I'm not claiming: it's a stylized visualization of one run, not a screen recording. My Korean prompt scored 0.29 against the 0.30 cutoff, so I re-routed an English brief. Jev's tier is "unvalidated" (a transport-checked verdict, not a calibrated guarantee), token numbers are estimates rather than billing, and routing is advisory: the host agent decides. Jev needs its own opt-in and sends data to a hosted API.

Full 60s trailer (with sound): x.com/JeYeong32741/status/2120893518470336551

Agensi: the jeo-skill router package, with setup guidance through Agensi MCP, is paid ($9.90/month or $95/year): agensi.io/skills/jeo-skill

The package is the router and setup materials, not a bundled copy of the catalog. The catalog itself is free and open source on GitHub (project-authored material is MIT, third-party skills keep their own licenses): github.com/akillness/jeo-skills

Feedback welcome from anyone running agents with big skill or tool libraries: do you route, prune, or just keep the list short?