Skills

A skill is a procedure you write once and the agent follows when the task calls for it. Skills work across Claude Code, Codex CLI, OpenCode and Cursor. Two related ways to extend an agent have their own pages: MCP servers add tools, and hooks run your commands automatically.

	Skill	MCP server	Hook
What it is	a folder with SKILL.md (and optional scripts)	a program that exposes tools and data the agent can call	a command the agent tool runs at a fixed point, such as after every edit
Who decides when it runs	the agent, when the task matches the description (or you, by name)	the agent, when it calls a tool	the tool itself, every time, whatever the model decides
In context	name and description always; the body only when used	every tool definition, on every turn	nothing, unless the hook returns output
Use it for	how you want a task done: benchmarking, releasing, publishing a trace	things the agent cannot do alone: current docs, a browser, GitHub, a simulation code, an HPC facility	rules that must always hold: format after edits, block dangerous commands, notify when done

Write a skill

A skill is a folder with a SKILL.md file:

---
name: benchmark
description: Measure performance before and after a change. Use when asked to profile, optimize or speed up code.
---

1. Run the baseline: `uv run python bench/run.py --label before`.
2. Profile the slow case, change one thing, keep the tests green.
3. Re-measure and report a before/after table.

The description decides when the agent loads the skill, so say when to use it, not only what it does.

Tool	Where skills live	Call one by name
Claude Code	.claude/skills/<name>/SKILL.md, ~/.claude/skills/	/name
Codex CLI	.agents/skills/, ~/.agents/skills/	$name
OpenCode	.opencode/skills/, also reads .claude/skills/ and .agents/skills/	/name
Cursor	.cursor/skills/, also reads .agents/skills/, .claude/skills/	/name

Paths change between versions; check your tool's documentation. Keeping skills in .agents/skills/ (with a .claude/skills link for Claude Code) gives one copy for every tool.

Example from this session: the demo replays were made with a publish-agent-trace skill. It archives a sanitized copy of the agent session in the repository and builds an HTML replay with claude-replay, so others can see how the code was written.

Where to find skills

• Genesis Skills: the Genesis Mission collection of skills for science and engineering, including HPC, AmSC, literature search, plasma simulation, scientific-data metadata and AI-system evaluation. It works with Claude Code, Codex, Cursor, OpenCode and Gemini CLI. Development, issues and pull requests happen on the GitHub upstream; see below to install it.
• anthropics/skills: Anthropic's examples, including document and design skills.
• openai/skills: skills for Codex.
• K-Dense-AI/claude-scientific-skills: a large collection for scientific work (bioinformatics, chemistry, data analysis).
• addyosmani/agent-skills: engineering skills, with setup notes for several tools.
• vercel-labs/skills: a CLI that installs skills into any of the supported agents.
• agentskills.io: the open SKILL.md format.

Install Genesis Skills

The repository groups skills by domain (skills/<domain>/<skill>/). Agents expect each skill directly in their skills folder, so its unpack.sh script copies them there:

git clone https://gitlab.osti.gov/genesis/genesis-skills.git
cd genesis-skills
./unpack.sh --list                                          # preview the skills
./unpack.sh --target ~/.claude/skills --mode copy           # Claude Code
./unpack.sh --target ~/.agents/skills --mode copy           # Codex, Cursor, OpenCode, Gemini CLI

Add a domain name, such as ./unpack.sh hpc-skills --target ..., to install one domain only. Use --mode symlink to keep the skills linked to the clone, so that git pull updates them. Skills with the same name already in the target folder are replaced. See the repository README for details.

Read a skill before you install it: it is instructions the agent will follow, and it can include scripts.