wuyoscar/jev-skill

⚡ Awesome Jev Skills – What It Is

Jev is a framework for “coding agents”: AI systems that decide what to do, classify options, and score outcomes. The jev‑skill repository is a curated collection of skills (small, reusable pieces of logic) that let a Jev‑powered agent perform concrete tasks such as:

• Designing and batching typed questions for the model (jev skill)
• Sorting and labeling records (jev‑triage)
• Finding and verifying source evidence (jev‑documents)
• Evaluating outputs against rubrics (jev‑eval)
• Selecting the next legal action in a UI or game (jev‑act)

The repo also indexes 67 community projects, 5 skill definitions, and 108 example scenarios (with 14 recorded input/output pairs) that demonstrate how to plug the skills into real applications—browsers, games, code‑review dashboards, music generation, search tools, and more.

---

📦 How to Use It

1. Install the skill set – copy a short “install prompt” into a coding‑agent interface (Codex, Claude Code, OpenCode, etc.). The prompt pulls the skill definitions from the repository.
2. Pick a skill – each skill lives in its own folder (skills/jev/, skills/jev‑triage/, …) with a one‑page description and a template for the agent to follow.
3. Run a scenario – the README lists concrete scenarios (e.g., “next browser action”, “urgency screening”, “claim‑to‑source check”). The agent fills the template, calls the Jev model, and receives a structured answer (choice, score, or “Noul” evidence).
4. Extend or contribute – the repo encourages adding new demos, projects, or scenario templates via the contribution guide.

---

🛠️ Key Components

Component	Purpose	Example Use‑Case
Skills (skills/…/SKILL.md)	Define a single, well‑scoped job for the agent (question design, triage, document lookup, evaluation, action selection).	jev‑act picks the next legal move in a Tetris game.
Scenarios (skills/**/references/scenarios.md)	Ready‑made prompts and expected inputs/outputs that illustrate the skill in action.	“Urgency screening” for incoming support tickets.
Projects / Apps (## Projects)	Community‑built demos that consume the skills (browser automation, code‑review dashboards, music composition, search re‑ranking, etc.).	Jev Ultrafast – a tiny LLM that types into a browser based on Jev decisions.
Local Model Adapters	Links to open‑source models that implement Jev’s “typed‑decision” API (e.g., OpenJev, Jevify, LitJev).	Running Jev on a laptop with a Qwen‑2.5‑1B model.
Tools & Resources	CLI utilities, evaluation suites, and research papers that help developers benchmark or extend Jev.	SemDecide for building JSON‑L pipelines, jev‑evaluation for adversarial testing.

---

🚀 Why It Matters

• Modular AI agents – By breaking agent behavior into tiny, typed decisions (Choice, Score, Noul), developers can reason about, test, and improve each step.
• Ecosystem hub – The repo aggregates dozens of community projects, model adapters, and evaluation tools, saving developers the hunt for compatible resources.
• Low‑code entry point – The “install prompt” lets non‑experts give a coding agent a ready‑made toolbox without writing any code themselves.

---

📚 Getting Started

1. Read docs/install.md for the exact one‑line prompt to paste into your agent.
2. Browse the Skills table, click a skill, and follow its short guide to try a scenario.
3. Explore the Projects table for full‑stack demos (browser automation, code review, music generation, etc.).
4. If you have a new use‑case, add a scenario or demo via the contribution guide.

---

TL;DR – jev‑skill is a community‑maintained toolbox that gives Jev‑based AI agents concrete, reusable abilities (triage, document lookup, evaluation, action selection) and points you at dozens of demos and open‑source models to try them out.