How coding agents pick up a skill, and how to call one yourself

What the agent loads, and when

1. At the start of a session the agent loads each skill's name and description, about 100 tokens a skill.
2. When your request matches a description, the agent reads that skill's SKILL.md. Anthropic advises keeping it under 5,000 tokens.
3. Files the skill points to (references, templates) load only when the agent opens them. A script runs through the shell, and only its output reaches the conversation.

That order is why a large library of skills stays cheap, and why the description carries all the weight: it is the only part the agent sees before it decides.

Calling a skill by name

In Claude Code, type a slash and the skill's name. Anything after the name reaches the skill as its arguments. A skill from a plugin carries the plugin's name first.

/frontend-design
/expo:expo-router add a settings tab

Codex runs a skill with $skill-name or lists them with /skills. Gemini CLI loads a skill through its own activate_skill step and asks you to approve it first.

When a skill never triggers

Open its SKILL.md and read the description. It has to say what the skill does and when to use it, in the words you would use when asking. "Helps with design" matches nothing; "Reviews UI code for accessibility and usability problems. Use when asked to review a page or check accessibility" matches the request you will actually type.

Claude Code shows the description in its skill listing and cuts it at 1,536 characters, so put the trigger words first.

Who may run it

Two frontmatter fields in Claude Code decide who starts a skill. disable-model-invocation: true leaves it to you, for skills with side effects such as a deploy. user-invocable: false leaves it to the agent, for background knowledge you never call by hand.

---
name: deploy-preview
description: Deploys the current branch to a preview URL. Use when asked to deploy or share a preview.
disable-model-invocation: true
---