> ## Documentation Index > Fetch the complete documentation index at: https://keystroke.ai/docs/llms.txt > Use this file to discover all available pages before exploring further. # Create skills > Write a SKILL.md with clear task guidance. A skill is a folder under `src/skills/` containing a `SKILL.md` and any supporting references. This page covers the format and the conventions that keep a skill useful. ## Create the folder Make a folder named for the skill (the folder name is how you'll [attach it to an agent](/docs/learn/skills/attach-skills-to-agents)) and add a `SKILL.md`: ``` src/skills/support/ SKILL.md references/ refund-policy.md ``` ## Write SKILL.md A `SKILL.md` follows the [Agent Skills specification](https://agentskills.io/specification): YAML frontmatter with a `name` and a `description`, then a focused body. ```md src/skills/support/SKILL.md theme={null} --- name: support description: Use when answering customer support questions about product policy, refunds, or tone. --- # Support skill Follow the refund policy in the product guide and keep replies concise. Escalate billing disputes over $500 to a human. ``` | Frontmatter | What it does | | ------------- | ------------------------------------------------------------------- | | `name` | The skill's name; match it to the folder name | | `description` | Intent-based routing rule for when the agent should load this skill | Write the `description` as a **Use when…** rule. Keystroke advertises attached skills to the agent as a compact catalog of names and descriptions, then the agent loads matching skills with `skill_view`. The description is the routing signal — describe the situation, not the file contents. ## Keep the body short, push detail to references Skills work by **progressive disclosure**: the agent loads the short `SKILL.md` first via `skill_view`, then loads deeper material only when the task needs it. Keep `SKILL.md` to the essentials and put long policies, examples, and edge cases in a `references/` folder: ``` src/skills/support/ SKILL.md # when to use it + the key rules references/ refund-policy.md # the full policy escalation-matrix.md # detailed procedures ``` Reference files are a convention, not a fixed structure; name and organize them however reads best. Everything in the skill folder is materialized into the agent's workspace at `/workspace/agent/skills/{slug}/`. After loading `SKILL.md`, the agent can open a packed reference with `skill_view({ name, file_path })`. References aren't the only kind of supporting material. The whole folder is copied verbatim, so a skill can also ship helper scripts the agent runs with its shell tools, or assets like templates and fixtures it reads: ``` src/skills/report/ SKILL.md references/format.md scripts/build-report.ts assets/template.html ``` Point the agent at them by relative path from `SKILL.md`. ## Next steps Attach a skill to an agent by folder name. What project skills are and when to use them. Reuse skills from external registries. Add static reference documents to a workspace.