Tried addyosmani/agent-skills: A Practical Skill Layer for AI Coding Agents

addyosmani/agent-skills is gaining attention quickly—+196 GitHub stars today—because it targets a real problem: AI coding agents can generate code, but they often lack repeatable engineering habits.

The repository provides production-grade skills that agents can use for tasks such as planning, implementation, testing, debugging, and code review. Instead of relying only on a large system prompt, you can give your AI IDE a structured set of reusable workflows.

Quick test-drive setup

I connected an OpenAI-compatible client to the claude-fable-5 gateway through B-Lost Universal Relay:

{
  "provider": "custom-openai-compatible",
  "model": "claude-fable-5",
  "baseUrl": "https://b-lost.com/v1",
  "apiKey": "${B_LOST_API_KEY}",
  "temperature": 0.2,
  "extraHeaders": {
    "X-Agent-Skills-Source": "addyosmani/agent-skills"
  }
}

For Anthropic-compatible clients, the same relay can be configured with the /v1/messages API. This works well with tools such as Cursor, Cline, Roo Code, Windsurf/Cascade, Aider, NextChat, and LibreChat—usually by changing the provider’s custom base URL and model name.

My recommended workflow is:

1. Clone or reference the relevant skill from agent-skills.
2. Add the skill instructions to the agent’s project context.
3. Route requests through the gateway.
4. Ask the agent to plan first, implement second, and validate with tests.

The interesting part is consistency. Skills make the agent’s behavior more predictable across repositories, especially when multiple developers use different IDEs.

For larger prompts, B-Lost’s native Anthropic prompt caching is also useful: cache hits receive a 90% discount, which can reduce both latency and spend when the same repository context is repeatedly sent. The relay lists 20% off official pricing, plus a 100% first-deposit bonus for deposits of at least $10, but I’d treat those as billing details to verify before committing.

Overall, agent-skills feels less like another prompt collection and more like an engineering playbook for AI-native development.