Agent Skills went from Anthropic experiment to open standard with 25k GitHub stars

A folder with a SKILL.md file is becoming the npm package of agent workflows.

Agent Skills , the open format originally developed by Anthropic, now has more than 25,000 GitHub stars. The Vercel skills CLI (npx skills) indexes tens of thousands of installable skills. Cursor, Copilot, Codex, and Gemini CLI all read the same shape.

I ship skills in client repos (blog writing, digest automation, SEO checklists). This is the infrastructure layer underneath those deliverables.

What a skill actually is

At minimum, a skill is a directory with SKILL.md:

---
name: my-skill
description: What this skill does and when to activate it. Include keywords agents can match.
---

## Instructions

Step-by-step guidance the agent follows when the skill is relevant.

The official specification defines required frontmatter:

Optional folders:

my-skill/
├── SKILL.md
├── scripts/       # Executable helpers
├── references/    # Longer docs loaded on demand
└── assets/        # Templates, images, fixtures

Agents use progressive disclosure: metadata loads at startup (~100 tokens per skill), full instructions load when activated, references load only when needed. That keeps context lean compared to dumping a 50-page playbook into every prompt.

Why the format won

The playbook mirrors MCP: build something useful internally, open-source the interface, let the ecosystem adopt it.

Skills solve a boring problem: agents do not know your team's procedures. MCP connects tools. Skills connect judgment: how we write blog posts, how we run migrations, how we review PRs.

Installing and sharing skills

The Vercel CLI is the default package manager:

npx skills add anthropics/skills
npx skills add addyosmani/agent-skills
npx skills add huggingface/skills

Useful flags:

Discovery also lives on skills.sh with leaderboards by installs and trending scores.

Popular packs worth skimming (not endorsements, just signal):

• anthropics/skills (reference implementations)
• addyosmani/agent-skills (engineering lifecycle)
• huggingface/skills (Hub and ML workflows)

Where skills live per agent

This repo uses .agents/skills/ for blog-writing, content-strategy, and digest automation. Same format, portable across clients.

Writing a skill that actually gets used

Bad skills read like generic blog posts. Good skills read like a senior engineer's checklist:

1. Description is the trigger. Include verbs and nouns your agent will see in user messages ("MDX blog post", "cover image", "no em dashes").
2. Front-load constraints. Hard rules before background essay.
3. Split long references. Keep SKILL.md under ~500 lines; move tables and examples to references/.
4. Ship verification steps. "Before shipping, confirm X" beats "be careful."
5. Version in git. Skills are code. Review them in PRs.

learn-agent-skills is a solid tutorial series if you want a guided first skill (README Wizard end to end).

Skills vs MCP vs AGENTS.md

Use all three. Skills are not a replacement for MCP any more than npm packages replaced HTTP.

What I would standardize this quarter

If you are rolling out agents to a team:

1. Pick one skills directory convention (.agents/skills/ travels well)
2. Add npx skills to onboarding docs
3. Require a skill for any workflow repeated more than twice per week
4. Lint descriptions the way you lint commit messages (vague descriptions never activate)

The 22k-star headline from newsletters is really infrastructure maturity. Agents are only as consistent as the procedures you package.

If you want help turning your internal playbooks into portable skills (without dumping secrets into prompts), book a free discovery call.