What are Agent Skills?
Agent Skills are folders of instructions, scripts and reference files that teach an agent how to do a particular job. Each skill centres on a SKILL.md file, and the agent loads a skill only when the task at hand calls for it.
How a skill is structured
A skill is a folder named after it. The YAML frontmatter of SKILL.md holds name and description fields, and the body holds step-by-step instructions. The folder can also carry scripts, templates or reference documents.
- In Claude Code, personal skills live in ~/.claude/skills/ and project skills in .claude/skills/.
- Codex supports skills too: project skills live in .agents/skills and personal ones in ~/.codex/skills.
- Plugins can bring skills with them too.
Why skills load gradually
At the start, only each skill’s name and short description go into the model’s context. When a task matches a skill, the model reads the full SKILL.md and, if needed, the files next to it. That keeps the context small even with dozens of skills installed, but a skill with a vague description won’t kick in at the right moment.
Skills, MCP, subagents and hooks
These four get mixed up a lot; each solves a different problem.
- Skill: know-how and scripts for a job; the model decides when to use it.
- MCP: adds new tools and connections to external services.
- Subagent: a helper agent that does work in a separate context.
- Hook: a command that runs on a given event every time, regardless of the model.
Writing a good skill
A skill is only as useful as its ability to trigger at the right time and stay short.
- In description, say clearly what the skill does and when to use it.
- Keep SKILL.md short and move long references into separate files.
- Turn repeated, exact steps into scripts so the model doesn’t rewrite them each time.
- Read other people’s skills before installing them; their scripts run on your machine.
Skills in AgentVera
AgentVera’s skill gallery brings the Claude Code and Codex skills on your computer into one list, and context profiles let you control how much skills add to the first request.
- Details → Automation → Skills: create, enable, disable, edit or copy Claude Code and Codex skills to another agent.
- Skills installed from the gallery have their files checked with SHA-256 and are scanned on your computer before installing; scripts are never made executable.
- Sharing your own skills with your organization or the public gallery is in the Pro and Team plans; public skills go through moderator approval.
- The Lean context profile turns off user plugins, hooks, skills and external MCP, cutting Claude Code’s first request from about 34.4k to 13.3k tokens.
FAQ
What’s the difference between a skill and MCP?
MCP gives an agent new tools and connections; a skill teaches it how to do a job. A skill can even describe how to use MCP tools along the way.
Do skills use tokens on every request?
Only their names and descriptions are in context on every request. The full content loads when the model decides to use the skill.
How do I get the agent to use a skill?
Usually you don’t need to do anything; the model loads a skill when the task matches its description. You can also mention the skill by name in your message and ask the agent to use it.
Is it safe to install someone else’s skill?
It depends on what’s inside. Skills contain instructions and scripts, and the scripts run on your computer. Read SKILL.md and the scripts before installing.