Agent Skills Collections: Teaching AI Assistants Your Team's Way of Working

Ever notice how AI assistants give great generic advice, but struggle with your team’s specific workflows? Agent Skills fix that. Here’s what I learned building them.

What’s an Agent Skill, Really?

OK, fancy name aside - an Agent Skill is just a folder with instructions and scripts that an AI can read and use.

Think of it like this: instead of repeatedly explaining to GitHub Copilot or Claude how your deployment process works, you write it down once in a structured way. Then the AI can reference it whenever needed.

.skills/
├── skill.md           # "What this skill does"
├── instructions.md      # "How to actually do it"
├── scripts/             # The actual code
└── examples/            # "Here's what good looks like"

It’s basically npm for AI capabilities. Which sounds obvious in hindsight, but took us way too long to figure out.

The Problem This Solves

You’ve definitely hit this before:

You: “Deploy my app to production”

AI: “Sure! Here are the general steps for deploying applications…” [proceeds to give you a tutorial you don’t need]

You: sighs and does it manually because explaining your CI/CD setup would take longer

With a deployment skill installed:

You: “Deploy my app to production”

AI: [reads your deployment skill] “Running your build script… pushing to Cloudflare… Done. Live at workers.dev/yourapp”

You: “Wait, that actually worked?”

The AI knows your team’s conventions, your security policies, your quirks. No more generic advice.

How It Actually Works

When you ask an AI to do something, it can now:

1. Search its available skills (“Do I have a ‘deployment’ skill?”)
2. Load the instructions (“Ah yes, here’s how this team deploys”)
3. Execute the actual work (run scripts, call APIs, etc.)
4. Give you real results

Here’s how it flows:

sequenceDiagram
    User->>Agent: "Deploy to production"
    Agent->>Skills Registry: Search for "deployment" skills
    Skills Registry->>Agent: Found "cloudflare-deploy" skill
    Agent->>Skill Folder: Load instructions.md
    Skill Folder->>Agent: Step-by-step deployment guide
    Agent->>Scripts: Execute deployment script
    Scripts->>Cloudflare API: Deploy application
    Cloudflare API->>Agent: Deployment successful
    Agent->>User: "Deployed to workers.dev/my-app"

Notice how informal this is? That’s intentional. Write it like you’d explain it to a teammate.

Where Skills Are Already Working

Vercel Agent Skills

A curated collection of modular skills for AI coding agents. Each skill is a packaged set of instructions and scripts that extend an agent’s capabilities by following the Agent Skills format. ([GitHub][1])

Cloudflare Skill (dmmulroy/cloudflare-skill)

A comprehensive Cloudflare platform reference skill for AI/agent workflows (e.g., OpenCode agents). Includes decision trees, product patterns, and detailed documentation covering Cloudflare services such as Workers, Pages, storage, networking, AI, and security. ([GitHub][1])

Next Started

My advice? Pick one thing you keep explaining to AI assistants that they get wrong. Write a skill for it.

Could be:

• How to run your test suite
• How to format your commit messages
• How to deploy to your specific cloud setup
• How to query your specific database schema

Document it once as a skill. Share it with your team. Watch the AI actually understand your workflow.

The format is still evolving, but the core idea - giving AI access to procedural knowledge - that’s here to stay.

Worth checking out:

• modelcontextprotocol.io - GitHub’s open standard
• anthropic.com/tool-use - Claude’s approach
• python.langchain.com - LangChain tools
• github.com/modelcontextprotocol/servers - Example skills

For the official spec and more examples: agentskills.io