Guide
What is a Claude skill?
A Claude skill is a SKILL.md file — markdown instructions with YAML frontmatter, plus optional scripts and resources — that the agent loads automatically when the skill’s description matches the task. It teaches the agent how to do something well and consistently, without you re-explaining it each time. The format is cross-agent: the same SKILL.md runs in Claude Code, Cursor, Codex, Gemini CLI, and Antigravity.
The whole point is selective loading. A skill announces itself with a short description; the agent only pulls in the full instructions when they are relevant, so your context window is not permanently spent on capabilities you are not using.
How they load
Three tiers, progressive disclosure
- 1
Tier 1 — Metadata (always loaded)
The name and description from the SKILL.md frontmatter, roughly 100 tokens, stay in context. The agent scans these to decide whether a skill is relevant. A weak description is the single biggest reason a skill never triggers.
- 2
Tier 2 — Instructions (loaded on trigger)
When the description matches the task, the agent loads the markdown body of the SKILL.md — the actual instructions — typically under 5K tokens. Once loaded, it stays in context across turns, so every line is a recurring token cost.
- 3
Tier 3 — Bundled files (loaded on demand)
Scripts, reference docs, and templates the skill ships are only read when the instructions call for them. This progressive disclosure keeps the agent efficient — it never pays for what it does not use.
What’s inside
A minimal SKILL.md
Only name and description are required by the spec. The description is the field that decides whether the skill ever triggers.
---
name: pdf-form-filler
description: Fill, flatten, and extract fields from PDF forms.
Use when the user mentions filling a PDF, AcroForm fields,
or merging data into a PDF template.
---
# PDF form filler
When filling a PDF form:
1. Inspect the AcroForm field names with the bundled script.
2. Map the user's data to those field names.
3. Fill, then flatten so values are not editable.
...For the full pattern — frontmatter fields, when to add scripts, how to write a description that triggers reliably — see the anatomy of a SKILL.md.
FAQ
Common questions
What is a Claude skill?
A Claude skill is a folder containing a SKILL.md file — markdown instructions with YAML frontmatter, plus optional scripts and resources — that the agent loads dynamically when the skill’s description matches the task at hand. It teaches the agent how to do something well and consistently. The format is cross-agent: the same SKILL.md runs in Claude Code, Cursor, Codex, Gemini CLI, and Antigravity.
What is in a SKILL.md file?
Two parts. YAML frontmatter between --- markers, where only name and description are required by the spec; and the markdown body with the instructions the agent follows. The description is the most important field — the agent decides whether to load the skill based almost entirely on it.
How is a skill different from just a prompt?
A prompt is one-off and lives in your message. A skill is reusable, versioned, and loaded automatically when relevant — you do not have to remember to paste it. It can also bundle scripts and reference files a prompt cannot.
How is a skill different from an MCP server?
A skill changes how the agent works (methodology) and loads into the current context. An MCP server changes what the agent can reach (databases, APIs, browsers). Most workflows use both. See skills vs subagents vs MCP vs plugins for the full picture.
Do Claude skills work in Cursor and Codex?
SKILL.md is a cross-agent standard, so skills nominally work across agents — but real support varies per skill and agent. On this site, each skill page lists only the agents where compatibility is verified, rather than assuming it.
Why does a skill sometimes not activate?
Almost always the description. The agent matches the task against the skill’s description; if it is vague or omits the literal phrases a user would type, nothing matches and the skill never loads. Editing the SKILL.md mid-session also will not take effect until you restart the agent.
Next: How to install a skill · How to create one · Best skills by use case