What Goes Into a Skill File
A skill is a Markdown file with two jobs. The frontmatter tells the agent what this file is and when it deserves context. The body gives the agent material to work with, ready to read the moment the file loads.
Two frontmatter fields carry the weight. name is the identifier the agent lists. description is the router: the agent reads it before every task and loads the file only when the description matches what you asked for. A vague description produces a skill that never loads. The loading mechanics are covered in the five layers of a real skill system.
The description does the routing. Write it like a trigger: name the topic and the moment to load the file. The body can run long. The description has one job and it runs on every single request.
What People Search For When They Want One
Autocomplete is a cheap demand check. I typed create agent skill into the suggest endpoint in November 2026 and it offered these completions:
| What Google completes | What it tells you |
|---|---|
| create agent skills | People build these in plural, one per topic |
| create agent skill claude | Claude users are the largest group reaching for the format |
| create agent skills github | Searchers want examples they can clone |
| create agent skill claude code | The demand sits inside coding workflows |
| create agent skill copilot | Copilot reads the same file inside its own app |
Google autocomplete probe for the exact phrase create agent skill, November 2026. Claude and Copilot in the same list tells you who is building: people already living in agent tools, reaching for a file they can version instead of another chat thread.
Three Ways to Get From a URL to a Skill
You have three routes to the same file. Each one trades control against setup, and the right pick depends on how many pages you intend to convert.
| Manual copy-paste | CLI crawler | Browser generator | |
|---|---|---|---|
| Setup | None | Clone a repo, install a runtime | None |
| Speed | Slow, once per page | Fast after the first run | Seconds, every time |
| Output | Whatever you paste | Everything the crawler walks | One clean page, capped for skill size |
| Best for | Pages you already know | A whole documentation site | A skill from any URL, today |
The browser route is the one I shipped. A crawler is heavy machinery when your job is one reference page. Copy-paste works until the third revision, when the formatting starts fighting you.
The Pipeline That Builds the File
Every generator performs the same five moves. The differences are where the work runs and what touches your content along the way.
- Paste. The tool normalizes the URL, scheme included, so a bare domain works.
- Fetch. One hop through a fetch worker. Public pages only, since the request carries no cookies and no login.
- Clean. Navigation, scripts, footers and ads come off before any conversion runs.
- Convert. The readable body becomes Markdown inside your browser.
- Wrap. Frontmatter goes on top: name from the page title, description from the meta description, source URL and date underneath.
Everything except the fetch runs in your browser. The worker returns the page and keeps no log of the URL. No model touches the content, which is why the build needs no key and no account.
Try it on the documentation page you rely on most: open Create Agent Skill, paste the URL, rename the skill if you want, and download the file. The whole loop takes seconds.
Write the Description Like a Trigger
The generator fills the description from the page meta description, which is a starting point rather than a finished line. Rewrite it as a condition: load this when the task involves the Acme API. Summarizing the page wastes the slot, because the page itself sits right below in the reference section.
Ten seconds spent here decides whether the skill loads on the tasks you care about. A description that names the site, the topic and the moment will beat a clever summary every time the agent has to choose.
Where the File Goes
Drop the folder into your agent's skills directory. For Claude Code that means .claude/skills/ inside the repository, or ~/.claude/skills/ when you want it available everywhere. The agent picks up new skills at the start of the next session.
Repository rules live in a different file. AGENTS.md sits at the root and loads every session; skills load when the description earns them. Most working setups keep both: rules in AGENTS.md, references as skills. The reasoning behind that split is in my AGENTS.md memory architecture.
When a Skill Is the Wrong Tool
If the task happens once, a prompt in the conversation is enough. If the content changes hourly, the skill goes stale the day after you build it. And if the goal is for crawlers to read your site rather than for an agent to load a reference, that is an llms.txt job: same audience of machines, different file.
Skills earn their folder when the reference stays stable, the task repeats, and the agent should reach for the file without being told twice.
FAQ
What is an agent skill?
An agent skill is a Markdown file with frontmatter that teaches an agent one topic. The name identifies it, the description tells the agent when to load it, and the body carries the reference material. Skills load on demand, which keeps the main prompt small.
How is a skill different from a prompt?
A prompt is text you paste for one conversation. A skill is a file the agent loads when the task matches its description, then drops when the task is done. Prompts compete for attention in every message. Skills wait in a folder until they earn their context.
Do I need an LLM or an API key to create one?
No. The generator on this site fetches the page, strips the navigation, converts the readable content and adds frontmatter, all in your browser. No model is involved and no key is needed.
Where does the skill file go?
In your agent's skills folder. For Claude Code that is.claude/skills/ inside the repository or ~/.claude/skills/ for your user account. Other agents read a skills directory next to their config file.
The Bottom Line
The web is full of pages your agent re-fetches and re-parses on every session. A skill file freezes the useful part once and puts it where the agent looks first.
Build one from the page you depend on most with the Create Agent Skill generator, then edit the description like it matters, because it does. The format works with Claude Code, Codex and anything else that reads SKILL.md.
If the workflow around the file is the part that breaks, that is what the free diagnostic is for.