Title: Agent Skills Glossary: SKILL.md, MCP, Hooks and More ยท ProSkills URL: https://proskills.dev/glossary Description: Plain definitions of the terms around agent skills: SKILL.md, frontmatter, progressive disclosure, AGENTS.md, CLAUDE.md, MCP, hooks, subagents, prompt injection, scan scores and llms.txt. --- Reference # Agent skills glossary 28 terms you meet when you write, install or check agent skills, each defined in a sentence or two. Link to any term with its #anchor. Agent skillSKILL.mdFrontmatternamedescriptionallowed-toolsReference fileBundled scriptProject and user scopeProgressive disclosureTriggerContext windowToken budgetAGENTS.mdCLAUDE.mdMCP (Model Context Protocol)HookSubagentSlash commandPluginSkill registryPrompt injectionHidden UnicodeSkill scannerScan scoreprofile.mdllms.txtAgent-skills discovery index ## The skill itself **Agent skill** A folder with a SKILL.md file that teaches an AI agent one task or body of knowledge. The agent loads it when a request matches its description. It is an open format read by Claude Code, Cursor, Codex and Gemini CLI. The long explainer is [what are agent skills?](https://proskills.dev/what-are-agent-skills) **SKILL.md** The one required file in a skill. It starts with YAML frontmatter holding a name and a description, followed by Markdown instructions for the agent. Every field and limit: [SKILL.md format reference](https://proskills.dev/skill-md-format). **Frontmatter** The YAML block at the very top of SKILL.md, between two lines of three dashes. The open specification requires name and description and allows license, compatibility, metadata and allowed-tools. **name** The skill's id in the frontmatter. It is 1 to 64 characters of lowercase letters, numbers and hyphens, and it must match the folder name. **description** The frontmatter field that says what the skill does and when to use it, up to 1,024 characters. It is always loaded, and it is the main thing the agent reads when deciding whether to use the skill. **allowed-tools** An optional, experimental frontmatter field listing tools the skill may use without asking. Support differs between agents, so check each tool's docs before relying on it. **Reference file** A file beside SKILL.md, often in a references/ folder, that the agent reads only when the body points to it. It keeps long tables and examples out of context until they are needed. **Bundled script** Code shipped in a skill folder, usually in scripts/, that the agent can run. It gives exact, repeatable steps, and it is the part of a skill to review most carefully before you trust it. **Project and user scope** Where a skill is installed. A project skill sits in the repo, for example .claude/skills/, and is shared with everyone who uses it. A user skill sits in your home folder, for example ~/.claude/skills/, and applies to all your projects. Paths for each agent: [how to install skills](https://proskills.dev/install-claude-code-skills). ## How skills load **Progressive disclosure** Loading a skill in stages. The agent always sees the name and description, reads the SKILL.md body when the skill triggers, and opens other files only when the task needs them. [Progressive disclosure for skills](https://proskills.dev/progressive-disclosure-skills) shows how to split one. **Trigger** The moment an agent decides to load a skill, by matching your request to the skill's description. In Claude Code you can also trigger a skill directly by typing /skill-name. When it does not happen: [why a skill is not triggering](https://proskills.dev/skill-not-triggering). **Context window** All the text a model can see at once: system prompt, project files, skills, tool results and the conversation. Everything loaded competes for the same space, and too much of it makes instructions easier to miss. **Token budget** How much context a skill costs. The spec guidance is about 100 tokens for the name and description, which are always loaded, and under 5,000 tokens and 500 lines for the body, which loads on trigger. Measure and trim: [reduce skill token usage](https://proskills.dev/reduce-skill-token-usage). ## Neighbours of a skill **AGENTS.md** A Markdown file at the root of a repo with instructions for coding agents: build commands, layout and conventions. It is an open format read by Codex, Cursor and many other agents, and it is loaded every session. **CLAUDE.md** Claude Code's project memory file, loaded in full at the start of every session. Claude Code treats it as context, not as enforced rules, so keep it short and move procedures into skills. Side by side: [skills vs CLAUDE.md vs MCP](https://proskills.dev/skills-vs-claude-md-vs-mcp). **MCP (Model Context Protocol)** An open protocol that connects an AI agent to tools and data through MCP servers, such as a database, an issue tracker or a browser. MCP gives the agent abilities; a skill gives it a procedure, and a skill can explain how to use an MCP server's tools. **Hook** A command the agent's harness runs on an event, such as before a tool call or after a file edit. It runs outside the model, so unlike a skill or CLAUDE.md it is enforced. Claude Code documents hooks; check other agents' docs for theirs. **Subagent** A separate agent with its own context window and instructions that the main agent hands a task to. In Claude Code, subagents are Markdown files in .claude/agents/. They keep long side tasks out of the main conversation. **Slash command** A saved prompt you run by typing /name. In Claude Code, custom commands have merged into skills: a file in .claude/commands/ and a skill in .claude/skills/ both create a slash command. **Plugin** In Claude Code, a package that bundles skills, commands, subagents, hooks and MCP servers so they install together from a plugin marketplace. A skill is one possible part of a plugin. ## Finding and trusting skills **Skill registry** A site or index that lists skills so people and tools can find and install them. Open directories such as skills.sh list community skills as published; curated libraries such as ProSkills review and scan each skill before listing it. **Prompt injection** Text that tries to make an agent ignore its user and follow someone else's instructions. In a skill the instructions are trusted by design, so installing a skill from an unknown author means running that author's instructions with your permissions. What to look for: [is this skill safe?](https://proskills.dev/is-this-skill-safe) **Hidden Unicode** Invisible characters, such as zero-width spaces, Unicode tag characters and bidirectional controls, that a model can read but a person reviewing the file cannot see. They can carry instructions past a human review. A box that reveals them: [hidden Unicode in skills](https://proskills.dev/hidden-unicode-prompt-injection). **Skill scanner** A tool that checks a skill's files without running them: spec errors, token cost, unclear instructions and risky patterns such as remote code execution or secret exfiltration. A scanner catches known patterns but cannot prove a skill is safe. The ProSkills [Skill Scanner](https://proskills.dev/scan) runs in your browser and also as `npx proskills-cli scan`. **Scan score** The ProSkills scanner's 0 to 100 rating of a skill: four pillars of 25 points each for spec, efficiency, clarity and security. Grades are A at 90 or more, B at 80, C at 70 and D at 60. Any critical security finding makes the grade F. Scores for real skills: [skill quality leaderboard](https://proskills.dev/skill-quality-leaderboard). **profile.md** A file ProSkills ships beside each SKILL.md that holds your details: brand, audience, stack or conventions. The skill reads it first, and skill updates never overwrite it. ## Machine-readable sites **llms.txt** A Markdown file at a site's root, /llms.txt, that summarises the site and links its key pages for language models. It is a proposed convention from llmstxt.org. ProSkills publishes one at [/llms.txt](https://proskills.dev/llms.txt). **Agent-skills discovery index** A JSON file at /.well-known/agent-skills/index.json that lists the skills a site offers and where to download each one, so a tool can install skills from a site URL. The open skills CLI reads it, which is how npx skills add https://proskills.dev works. ## Sources Spec terms follow the [Agent Skills specification](https://agentskills.io/specification). Claude Code terms (CLAUDE.md, hooks, subagents, slash commands, plugins) follow [Claude Code's documentation](https://code.claude.com/docs/en/skills). Scan score and profile.md describe ProSkills' own tools. FAQ ## Questions **What is the difference between a skill and a plugin?** A skill is one folder with a SKILL.md. A Claude Code plugin is a package that can hold several skills plus commands, subagents, hooks and MCP servers, installed together. **Is SKILL.md the same as AGENTS.md?** No. AGENTS.md is one file of project facts loaded every session. SKILL.md defines a skill whose body loads only when a task matches its description. **What does a scan score of 90 mean?** On the ProSkills scanner, 90 or more is grade A: the skill lost at most 10 points across spec, efficiency, clarity and security, and has no critical finding. It does not prove the skill is safe. **Who maintains this glossary?** ProSkills (proskills.dev), a curated library of scanned agent skills. Definitions follow the Agent Skills specification and each agent's own documentation. Related ## Keep reading - [**Skill Scanner** Score any skill in your browser. Nothing is uploaded.](https://proskills.dev/scan) - [**SKILL.md format reference** Every field, its limit, the folder layout and the errors people hit.](https://proskills.dev/skill-md-format) - [**What are agent skills?** The definition, how a skill loads, what SKILL.md holds, and where to get one.](https://proskills.dev/what-are-agent-skills) - [**SKILL.md format reference** Every field, its limit, the folder layout and the errors people hit.](https://proskills.dev/skill-md-format) - [**All guides** Every guide, grouped by what you are trying to do.](https://proskills.dev/guides) Free tool ## Scan your own skill Paste a SKILL.md or drop a folder. It is scored in your browser and nothing is uploaded. [Open the Skill Scanner](https://proskills.dev/scan) Want a skill that already passes? Start with a free one.