Skip to main content
Once skills are installed, you invoke them by typing their name as a slash command in Claude Code. Claude loads the skill’s SKILL.md as context and follows its workflow automatically — no additional prompting needed.

​
Invoke a skill

1

Open Claude Code

Start a new Claude Code session in any project directory. Skills are available globally regardless of which project you have open.
2

Type the slash command

Type the skill name as a slash command and press Enter. For example:
/deep-research
Claude will prompt you for any information it needs — the research topic, the target audience, or the specific question to answer — and then begin the structured workflow.
3

Claude runs the workflow

When you invoke a skill, Claude loads the skill’s SKILL.md file as context. That file contains:
  • Numbered workflow phases with pointers to reference files for each phase
  • Golden rules that constrain Claude’s behaviour throughout
  • A self-review checklist Claude runs before delivering output
Claude works through each phase in order, loading additional reference files as needed. You do not need to guide it through the steps.
4

Review the self-review checklist

Before delivering final output, every skill instructs Claude to run through a checklist of quality criteria specific to that skill. If any item fails, Claude corrects the issue before continuing.This feedback loop is what makes skills reliable across sessions — errors are caught and fixed automatically rather than passed to you.

​
Example skills to try first

/deep-research is a good first skill. It runs parallel research agents across multiple dimensions (WHO, WHAT, HOW, WHERE, WHEN, WHY), applies source quality tiers, and produces a structured report with honest coverage gaps noted.
/deep-research
When prompted, describe the topic you want researched. For example: “the current state of open-source LLM fine-tuning tooling”.

​
How context loading works

Skills use progressive disclosure to stay efficient:
  1. Only SKILL.md loads when you invoke the slash command. It’s intentionally short (~50–100 lines) — a table of contents, not a manual.
  2. When Claude reaches a workflow phase that needs detailed guidance, it reads the relevant reference file for that phase.
  3. Reference files contain the templates, code patterns, and domain knowledge for their specific phase.
This means Claude’s context window stays focused on what it needs right now, rather than loading everything upfront.
All domain knowledge for a skill lives inside its directory. Claude does not rely on external docs, chat history, or assumptions — everything it needs is in the skill files.

​
Next steps

Explore the full skill library to find workflows for your use case: