AI Skill
Also known as: Claude Skill, Anthropic Skill, Markdown skill pack
Quick definition
An AI Skill is a markdown-formatted documentation file (SKILL.md) that teaches an AI agent platform-specific or domain-specific reasoning rules. Anthropic's Skills format pairs the document with structured frontmatter; the AI client (Claude Desktop, Claude Code, Antigravity) loads skills at session start and uses them as in-context guidance for tool calls.
Contents
What is an AI Skill?
An AI Skill is a markdown-formatted documentation file (SKILL.md by convention) designed to teach an AI agent platform-specific or domain-specific knowledge — the rules, constraints, conventions, and best practices the agent should reason within when performing tasks. Skills emerged as a formalization of the 'system prompt' pattern: instead of cramming domain knowledge into one giant prompt, package it as named, reusable, version-able markdown documents that AI clients load at session start.
Anthropic published the Skills format spec in late 2024 alongside Claude Desktop's skill-loading capability. A SKILL.md typically includes YAML frontmatter (name, description, version, scope) followed by markdown body covering: domain context, important constraints, common patterns, edge cases, and concrete examples. The AI client parses these at startup and injects relevant skills into the model's context when the user's task matches the skill's scope.
Skills vs MCP — different layers of AI agent integration
MCP (Model Context Protocol) and Skills solve adjacent problems. MCP is the runtime layer — it gives the model access to tools the agent can call (publish_post, schedule_post, get_analytics). Skills are the reasoning layer — they teach the model how to think about a domain before deciding which tools to call. Use both: MCP for capabilities, Skills for context.
Concretely: CodivUpload publishes both. The MCP server (`npx codivupload-mcp`) exposes ~25 tools to AI clients. The Skills npm package (`npx codivupload-skills`) installs 7 markdown skill files covering platform-specific reasoning (Instagram constraints, TikTok flags, YouTube BYOP setup, X long-form rules, etc.). The agent loads the skills at session start, then uses MCP tools informed by the loaded context. Without skills, an agent might call MCP correctly but make platform-specific mistakes (forget tiktok_privacy_level, miss YouTube tag length rules). With skills, the same MCP tool call produces correctly-configured posts.
What goes in a SKILL.md file
Five sections most skills include. (1) Frontmatter — YAML block at the top with name, description, version, when_to_invoke. (2) Domain context — high-level overview of what this skill covers and why it matters. (3) Critical constraints — non-negotiable rules the agent must respect (e.g., 'TikTok defaults to private — always set privacy_level: 0 for public posts'). (4) Common patterns — recipes for typical operations (cross-posting, scheduling, error handling). (5) Edge cases and gotchas — platform-specific mistakes to avoid.
The goal is concise, actionable knowledge — not a textbook. Most skill files are 500-2,000 words; longer skills cap their effectiveness because the model loses focus across too much context.
Where to find AI Skills
Anthropic maintains a public skills registry and several SaaS products publish their own skills. Common installation patterns: npm install + CLI installer that copies SKILL.md files into ~/.claude/skills/ (for Claude Desktop and Code), into .agent/skills/ for Antigravity, into .cursor/rules/ for Cursor. CodivUpload's installer (`npx codivupload-skills`) supports all six client paths via --client flag.
Install CodivUpload's AI Skills for Claude Desktop
bash
# Install all 7 skill packs (generic, instagram, youtube, tiktok, x,
# facebook-linkedin, agency) into Claude Desktop's skill directory.
npx codivupload-skills --client claude
# Or for Antigravity (Google's AI client) — installs into .agent/skills/
npx codivupload-skills --client antigravity
# Or for Cursor IDE — installs into .cursor/rules/
npx codivupload-skills --client cursor
# After install, restart your AI client. Next time you give it a task
# that matches a skill's scope, the skill auto-loads into context. The
# agent then reasons about platform constraints (TikTok privacy, YouTube
# tags, Instagram media types) without you having to prompt for them.
# Skills + MCP together:
# - Install skills: npx codivupload-skills --client claude
# - Add MCP server: claude mcp add codivupload \
# -e CODIVUPLOAD_API_KEY=cdv_... -- npx -y codivupload-mcp
# - Now the agent has reasoning context (skills) + runtime tools (MCP)Common pitfalls
- ×Loading too many irrelevant skills — exhausts the model's context window without payoff
- ×Skills with stale platform information — re-publish quarterly or whenever platforms change
- ×Treating skills as full documentation — skills are concise reasoning packs, not exhaustive docs
- ×Forgetting the AI client restart — skill changes require client restart in most implementations
Tips
- ✓Pair skills with MCP — capabilities + context together is the full integration
- ✓Version skills like code — semantic versioning, update logs, deprecation paths
- ✓Test skill effectiveness with concrete prompts — does the agent now handle the tricky case correctly?
- ✓Publish skills via npm or PyPI — easy install, easy update, automatic distribution
Frequently asked questions
How is a Skill different from a system prompt?+
Skills are named, reusable, version-controlled markdown files loaded at session start. System prompts are typically inline strings the developer includes in API requests. Skills work better for complex domains because they're separable, shareable, and discoverable; system prompts work for one-off use cases.
Which AI clients support Skills?+
Claude Desktop and Claude Code (Anthropic), Antigravity (Google's Gemini client). Cursor IDE has 'Rules' which are skill-equivalent. ChatGPT Pro Connectors support 'Custom GPT instructions' which serve a similar purpose. The exact format varies but the underlying pattern (markdown context loaded at session start) is consistent.
Can I write my own AI Skills?+
Yes. Anthropic's Skills format spec is public; create a SKILL.md with proper frontmatter and markdown body, place it in your AI client's skill directory, restart the client. For npm-distributable skills, package the SKILL.md files in a CLI installer that copies them to the right path per client.
How long should a Skill be?+
500-2,000 words for most skills. Longer skills cap their effectiveness — models lose focus across too much context. Concise, actionable rules outperform exhaustive documentation.
Are Skills part of MCP?+
No, they're complementary layers. MCP is the runtime tool-calling protocol. Skills are documentation that teaches the agent how to reason about a domain. Use both: MCP for what tools exist; Skills for how to use them correctly.
Install CodivUpload's 7 AI Skills in 60 seconds
npx codivupload-skills --client claude (or antigravity, cursor, zed, chatgpt). Adds platform-aware reasoning to your AI agent. Pair with the MCP server for the full integration.
See AI Skills setup guideRead next
Related glossary terms