Overview & result
This Pi extension removes the full Agent Skills catalog from each model request. Jev scores every enabled skill against the current task; Pi loads selected SKILL.md files only when needed.
From the original post
A Pi extension that keeps the Agent Skills catalog out of model requests and replaces it with one ranking tool
How Jev fits in the loop
- Keep the skill catalog locally but omit its large generated block from model prompts.
- Send the task as shared state and each skill description as a separate Jev Score question.
- Sort scored skills, load the strongest instructions and allow named retrieval through `skill_load`.
How to reproduce
- Install with `pi install git:github.com/safzanpirani/pi-jev-skill-picker` and configure a TypeSafe API key.
- Call `skill_search` with a concrete task; inspect ranked skills and their scores.
- Compare prompt usage with and without the catalog-stripping extension.
Why this build matters
Jev acts as a tool-selection layer deciding which instructions a coding agent needs, reducing repeated prompt payloads.
Limitations
- The 87.9% saving is author-reported for one captured prompt, not a general benchmark.
- If Jev is unavailable, lexical fallback returns metadata instead of full skill instructions.