COMMUNITY OBSERVEDMetric claims: AUTHOR REPORTEDSOURCE AUDITED⚑ Jev SkillsπŸ€– Jev Agent

Lomesh Dutta Skill Router

A Claude Code session-start hook uses Jev to rank installed skills against the user's goal.

Overview & Result

Lomesh Dutta built an open-source Claude Code hook that sends the session goal and installed-skill catalog to Jev, then recommends a matching skill, a search for a missing skill, or no skill.

How Jev fits in the loop

  1. Claude Code captures a session goal and enumerates installed skills.
  2. Jev ranks the bounded skill list with probabilities and confidence.
  3. Claude Code loads the suggested skill or searches skills.sh only when no installed skill fits.

How to reproduce

  1. Read the public README and install Python 3.12, uv and a TypeSafe API key.
  2. Follow the repository's uv sync and install-hook instructions, then start a Claude Code session.
  3. Run the repository evaluation harness if comparing its author-reported results.

Why this build matters

Makes skill discovery an explicit, inspectable decision rather than leaving many installed skills unused.

Limitations

  • The hook currently targets Claude Code; other agent hosts are not promised.
  • Repository evaluation numbers depend on its skill catalog and author-run tests.

Patterns