wuyoscar/jev-skill

What it solves

It provides a collection of specialized skills, workflows, and demos for coding agents to improve their decision-making capabilities. Instead of relying on general LLM reasoning for every step, it allows agents to use "Jev" (a system that chooses, classifies, and scores) to handle specific tasks like tool routing, triage, and evidence checking more reliably.

How it works

The project offers a set of "skills" that can be installed into coding agents (such as Codex, Claude Code, or OpenCode). These skills act as specialized modules for different types of decisions:

• jev: Designing questions and batching calls.
• jev-triage: Sorting and labeling records (e.g., support queue routing).
• jev-documents: Finding and verifying source evidence.
• jev-eval: Checking outputs against a rubric.
• jev-act: Choosing the next legal action (e.g., browser steps or game moves).

Agents can use these skills via an API (through OpenRouter or TypeSafe) or via a simulation mode where the agent itself handles the decision logic.

Who it’s for

Developers building or using AI coding agents who want to add structured decision-making, classification, and scoring capabilities to their agentic workflows.

Highlights

• Diverse Skill Set: Includes dedicated modules for triage, document evidence, evaluation, and action selection.
• Extensive Ecosystem: Features a catalog of 57 projects, 108 scenarios, and various community-built apps ranging from browser automation to music composition.
• Agent-First Installation: Designed to be installed by the coding agent itself via specific prompts.
• Flexible Backend: Supports official TypeSafe services, OpenRouter, or local model simulations.