Overview
Agent-Skills, created by Addy Osmani (Google Chrome engineering lead and author of *Learning Web Performance*), went trending on GitHub this week. The premise: today's AI coding agents behave like smart interns — they can write code but lack the engineering workflow of senior developers. They skip specs, ship entire features at once, skip tests, skip review, and push straight to main. The problem is not intelligence; it is missing process.
Core Design: 8 Slash Commands, 24 Skills
The toolkit defines eight slash commands, each corresponding to a stage of the software development lifecycle:
| Command | Stage | Principle |
|---------|-------|-----------|
| /spec | Definition | Spec before code |
| /plan | Planning | Small, atomic tasks |
| /build | Implementation | One slice at a time |
| /test | Verification | Tests are proof |
| /review | Review | Improve code health |
| /webperf | Performance | Measure before you optimize |
| /code-simplify | Simplification | Clarity over cleverness |
| /ship | Release | Faster is safer |
Behind these commands sit 24 Markdown skills describing what to do, what not to do, and how to evaluate quality at each stage.
Key Design: Automatic Skill Activation
Users do not invoke individual skills manually. Typing a slash command triggers the system to activate relevant skills automatically. For example, /build activates test-driven-development, code-review-and-quality, and conditionally api-and-interface-design or frontend-ui-engineering depending on context.
Compared to obra/superpowers, which is a heavier methodology-and-framework package, Agent-Skills is lighter and modular — each command can be adopted independently.
/build auto: One Approval, Autonomous Execution
The most notable mode is /build auto. Given a spec, it:
1. Generates an implementation plan as atomic tasks 2. Implements each task test-first 3. Commits each task separately 4. Pauses on failure or risky steps
Crucially, this removes inter-step human gating, not verification. Traditional agent workflows pause for confirmation after every task; /build auto shifts the human's role to approving the plan once, then letting the agent run autonomously with continuous test-based validation.
This aligns with the RPI workflow direction (one-time approval plus automated verification) and elevates the human approval granularity from per-task to per-plan.
Why It Matters
A GPT-4-class model given "implement this feature" will write no spec, ship the whole feature, skip tests, skip review, and push. It knows how to do each step — no one told it the step is required now. Agent-Skills encodes the answer to "what should be done next" as explicit instructions. The same applies to human junior engineers: a checklist immediately improves output quality, and the same holds for agents.
Author Context
Osmani's decades of web-performance expertise directly shape /webperf — the principle "measure before you optimize" is especially important for agents, since LLMs natively prefer giving answers over measuring first.
A Larger Picture
Read together with recent trending projects, a three-layer agent operating system is emerging:
- State layer — cloudflare/computer (give an agent a computer)
- Control layer — huangruiteng/loopx (give an agent a scheduler)
- Skills layer — addyosmani/agent-skills (give an agent a manual)
---
*Project: addyosmani/agent-skills · JavaScript/Markdown · 203 stars today · Author: Addy Osmani (Google Chrome team)*