5 min

Skills — recipes your assistant loads on demand

A folder of markdown that teaches your CLI a procedure. Auto-triggered by description, model-loaded as needed.

A skill is a folder of markdown (and optional helper scripts) that teaches your coding CLI a procedure — how to do something. The assistant reads the skill's description on every turn; when your request matches, it pulls the full instructions into context and follows them. Skills are the answer to "I keep explaining the same workflow over and over."

The mental model

• A skill lives in a folder: SKILL.md + optional scripts.
• The frontmatter has a description: — one sentence the model reads on every turn.
• When your request matches that sentence, the assistant loads the body and follows the procedure.
• Skills compose: one can call another. Many ship with helper scripts the assistant runs.

Anatomy of a SKILL.md

---
name: git-prune
description: Prune merged git worktrees and branches, clean
  remote tracking refs, and pull latest main.
---

# Git Prune

When the user asks to clean up worktrees or branches:

1. Run `git worktree list` and identify merged branches.
2. For each merged worktree: `git worktree remove <path>`.
3. `git fetch --prune` to drop deleted remote refs.
4. `git checkout main && git pull --ff-only`.

Never delete unmerged work without confirming.

Skills people actually keep

How skills relate to the other primitives

When to write your own

• You've explained the same procedure 3+ times.
• The right steps depend on a context you keep forgetting to mention (e.g. "we use SQLAlchemy, not Django ORM").
• The task has a checklist of preconditions or guardrails you want enforced every time.
• You want a workflow available across every session, every repo, with no copy-paste.