building-skills-from-patterns（从模式中培养技能）

Building Skills From Patterns

Skills are reusable SKILL.md files. This meta-skill tells the agent to promote repeated muscle memory into a named skill: research once, encode the workflow, reuse forever.

When to trigger

• The user has asked for the same sequence three or more times (e.g. “always run lint then tsc then test before commit”).
• The agent notices it is re-deriving the same steps on every task in this repo (e.g. “how we deploy preview branches”).
• A correction sounds like a policy (“never use raw SQL here — always the repository layer”) — pair with suggesting-cursor-rules if it should be always-on; use a skill if it is a procedure with steps.

Workflow

1. Name the pattern

Choose a short slug (lowercase, hyphens): verifying-api-before-merge, releasing-mobile-build, etc.

2. Draft SKILL.md

Create .cursor/skills/<slug>/SKILL.md (or in this repo’s pattern, copy from resources/<slug>/SKILL.md when contributing upstream).

Frontmatter:

---
name: <slug>
description: One line: what it does and when to use it. Ends with a clear trigger.
user-invocable: true   # optional, if the user should be able to invoke by name
---

Body sections (keep lean):

1. Title — human-readable.
2. When to use — bullets.
3. Steps — numbered, imperative, tool names where useful (npm, gh, MCP tools).
4. Notes — edge cases, safety, when not to use.

Match the tone of other skills in the repo: concrete commands, no filler.

3. Validate

• Description is specific enough for Cursor to match the skill when the user describes the task.
• Steps are executable by an agent without guessing repo layout (or say “detect package manager from lockfile”).
• No secrets or machine-specific paths.

4. Point the user to it

Tell the user where the file lives and that the agent will pick it up on the next chat in that workspace.

Relationship to rules and hooks

If the pattern is “every time I save, run X,” suggest a hook instead. If it is “when I ask to ship,” keep it as a skill.

Notes

• Prefer one skill per workflow — avoid megaskills that try to cover every situation.
• Update an existing skill instead of adding a duplicate if the workflow evolves.