How to Use Claude Skills, Step by Step
Published May 30, 2026
The fastest way to use a Claude skill is to drop its SKILL.md folder into your agent's skills directory, then describe the task in plain language so the agent triggers it. For Claude Code that directory is ~/.claude/skills or a project's .claude/skills; for Codex it is .agents/skills; OpenCode reads its own plus both of those. No command memorization required, the agent matches your request to the right skill on its own.
This guide covers all three agents on purpose. Most tutorials that rank for this query are Claude-only, but skills are portable, and the cross-agent detail is exactly where teams get tripped up.
A note on triggering
Skills are matched by their descriptions, not invoked by name. That means a vague request can miss a skill that would have helped. If a skill is not firing, make the request more specific, or check that its description clearly covers the job you are asking for.
Where to go next
- Not sure what skills are yet? Read what are Claude skills.
- Browse the curated skills marketplace to find your first install.
- For OpenAI Codex users, see the best Codex skills.
The steps
- 1. Find a skill worth installing Start from a curated source rather than a raw search. Pick a skill whose description maps to a job your team actually does, and check its cross-agent compatibility note before you commit.
- 2. Put the SKILL.md folder in the right directory For Claude Code, place the skill folder in ~/.claude/skills for personal use or .claude/skills inside a project to share it with the repo. For Codex, use .agents/skills. For OpenCode, use its skills directory, or reuse .claude/skills or .agents/skills, which it also reads.
- 3. Confirm the frontmatter is valid Every skill needs a name and a description in its frontmatter. Codex in particular requires both. If a skill fails to load, a missing or malformed description is the usual cause.
- 4. Trigger the skill by describing the task You do not call a skill by a command in most cases. You describe the task, and the agent matches your request against the skill descriptions and loads the right one. Clear, specific requests trigger more reliably.
- 5. Verify it actually ran Check that the agent used the skill's process, for example by looking for the scripts it ran or the structured output it produced. If a Claude-specific feature like a hook did not fire on Codex or OpenCode, that is the expected cross-agent caveat.
Skills mentioned
-
Agentic SEO Skill
Top pickDeterministic, LLM-first SEO and GEO audits for sites, blogs, and repos, with script-backed evidence and prioritized, confidence-labeled fixes.
- Claude Code: Works
- Codex: Works
- OpenCode: Unverified