AI Skills for Coding Agents¶
AI Skills are modular, reusable capabilities that turn general-purpose AI agents into domain specialists.
A skill is not a prompt. It is a structured package: instructions + optional scripts + reference material + output templates. The agent discovers skills automatically, loads them on demand, and follows their workflows.
Skills work as an open standard across Claude Code, GitHub Copilot, Cursor, Devin Desktop (formerly Windsurf), Codex CLI, and other agents that accept SKILL.md files.
Sections¶
| File | Topics |
|---|---|
| What Is a Skill | Definition, SKILL.md anatomy, frontmatter, instructions, real examples |
| How Agents Load Skills | Progressive disclosure, 3-level loading, discovery, intent matching |
| Skill Packaging | Folder layout, scripts, references, templates, versioning |
| Orchestration & Workflows | Composing skills into chains, context handoff, failure handling |
| Evaluation & Security | Eval harnesses, quality metrics, security audit, governance |
| Cross-Agent Compatibility | Claude Code, Copilot, Cursor, Devin Desktop (formerly Windsurf), Gemini CLI, Kiro — setup and portability |
| Claude Code Best Practices | .claude/ folder anatomy, CLAUDE.md, settings.json, commands, rules |
| Claude Code Hooks & Agents | Lifecycle hooks, agent personas, subagents, three-layer config system |
| Claude Code Workflow Patterns | Verify work, explore-plan-code, prompting, session management, failure patterns |
| Claude Code Advanced Config | Sandbox, plugins, MCP servers, non-interactive mode, enterprise settings |
| Model Context Protocol (MCP) | Protocol architecture, primitives, transport, security, popular servers, custom servers, skills integration |
| AGENTS.md Standard | Universal agent instructions, anatomy, effective patterns, anti-patterns, cross-tool compatibility, monorepo scoping |
| SKILL.md Playground | Hands-on exercises: write frontmatter, workflows, anti-rationalization tables, red flags, verification, folder layout |
| AGENTS.md Playground | Hands-on exercises: mission statement, toolchain registry, judgment boundaries, escalation, monorepo scoping |
| Claude Code Project Structure | Reference layout, key components, best practices, getting started, development tips |
| SKILL.md Universal Standard | Universal spec summary, anatomy, progressive disclosure, frontmatter, compatibility paths, implementation checklist |
| Building Skills for Claude: Practical Playbook | Condensed essentials from the full Claude skills guide: planning, frontmatter, testing, iteration, distribution |
| Skills Troubleshooting & Checklists | Pre-build and rollout checklists, failure diagnosis, trigger tuning, MCP debugging, proven workflow patterns |
| Claude Code Commands Reference | All built-in slash commands, keyboard shortcuts, CLI flags, common workflows |
Why Skills, Not Prompts¶
| Prompt | Skill | |
|---|---|---|
| Scope | one conversation | reusable across sessions/projects |
| Structure | free-form text | YAML frontmatter + markdown instructions + bundled files |
| Loading | always in context | loaded on demand (progressive disclosure) |
| Execution | LLM reasoning only | LLM reasoning + deterministic scripts |
| Testing | manual checks | structured eval harness |
| Portability | copy-paste | open standard across agents |