Skills

AI Navigate Original / 4/27/2026

💬 OpinionDeveloper Stack & InfrastructureTools & Practical Usage
共有:

Key Points

  • Skills extend an agent's "way of doing things" for you
  • Persist conventions/rules; AI auto-follows, killing re-teaching fatigue
  • One Markdown file; auto-activate or explicit call; share via Git
  • 1 Skill = 1 concern, ~10/project, no confidential info

What Are Skills

Skills are a mechanism to extend an AI coding agent's "way of doing things" for yourself. Register project-specific conventions, repetitive tasks, and internal rules as a Skill, and the AI always follows them. Similar features exist in Claude Code, Codex, Cursor (this article centers on Claude Skills, organizing the common idea).

The Problem Skills Solve

Do you write the same preamble to AI every time?

  • "Commit messages in English, Conventional Commits format"
  • "Tests with Vitest, in a describe/it hierarchy"
  • "When refactoring, first list the impact range"

Persisting these as a Skill makes the AI auto-follow next time. "Re-teaching fatigue" disappears.

Skill Structure

A Skill can basically be defined in one Markdown file:

---
name: e2e-test-writer
description: |
  Style guide for writing E2E tests with Vitest + Playwright.
  File naming, describe hierarchy, data cleanup.
---

# Naming
- 1 feature 1 file
- File name: feature-name.e2e.ts

# describe hierarchy
- Top: feature name
- 2nd: user-operation sequence
...

Usage (Claude Code)

1. Create a skills/ Directory in the Project

your-project/
  .claude/
    skills/
      e2e-test-writer.md
      api-error-handler.md
      release-note.md

2. Auto-Activate or Explicit Call

When Claude judges "this seems to fall under a Skill," it auto-loads the Skill content. Or call explicitly with /skills e2e-test-writer.

3. Share with the Team

Skills are version-managed in Git. The whole team using the same Skill unifies AI behavior.

Typical Skill Examples

Language / Framework Standards

  • "TypeScript strict, no any"
  • "React only function components + Hooks"
  • "Python type hints mandatory"

Business Flows

  • "DB migrations with expand-migrate-contract"
  • "Release notes in Keep a Changelog format"
  • "Security fixes: no public PR"

Custom-Tool Integration

  • "How to call the internal ID-issuance API"
  • "How to edit a special CI config"

Skills Equivalents in Other Tools

ToolSkills equivalent
Claude Code.claude/skills/*.md
Cursor.cursor/rules/*.mdc
GitHub Copilot.github/copilot-instructions.md
Codex CLIAGENTS.md (project root)

Order to Make Skills

  1. None initially: work with AI without Skills at first
  2. Note what you say every time: "saying this again" is a Skill candidate
  3. 1 Skill = 1 concern: split small over a mega Skill
  4. Maintain periodically: delete unused Skills. Keep them up to date

Cautions

  • Too many Skills confuse the AI: ~10 per project is the upper guide
  • CLAUDE.md / Cursor Rules may suffice over a Skill: one root file is OK if it affects the whole project
  • Don't put confidential info: Skills go into Git; secret keys etc. are strictly forbidden

Summary

Leveraging Skills makes AI behave like "your own dedicated colleague." It remembers what you taught, so productivity grows exponentially. "It's a pain to explain again" is the sign to make a Skill.