CrewAI Skills: Reusable Procedures and SKILL.md for Agent Teams

Skills vs Knowledge

Knowledge teaches an agent facts; skills teach procedures. A "Refund Policy" PDF is knowledge; "How to escalate a P0 incident" with steps and tools is a skill.

SKILL.md Format

A skill lives in a folder with a SKILL.md at the root. The file's frontmatter declares the skill's name and triggers; the body is the instruction text injected at activation.

Discovery & Activation

At startup, call discover_skills() to enumerate available skills. When an agent needs a specific skill, call activate_skill(name) — the body is injected into the prompt and associated tools/knowledge are wired up.

from crewai.skills import discover_skills, activate_skill

@before_kickoff
def setup(self, inputs):
    skills = {s.name: s for s in discover_skills()}
    if inputs.get("priority") == "P0":
        activate_skill("escalation_p0")
    return inputs

Crew-Level Skills

Wire skills that every agent in the crew shares — useful for company-wide policies (tone of voice, brand voice, compliance rules) — by activating them in the @before_kickoff hook.

Skills + Tools Together

A skill can declare tools it depends on. When activated, those tools become available to the agent automatically. This pattern keeps prompts focused: only enable what's needed for the task at hand.

RAG indexes behind knowledge-heavy skills

Skills that bundle document retrieval still rely on the global vector client configured at process start. Call set_rag_config() before crews load knowledge sources, and use clear_rag_config() in test teardown so parallel suites do not fight over the same on-disk index.

Best Practices

• ✓Keep SKILL.md bodies short (1 page or less) — long bodies inflate every prompt that activates them.
• ✓Use progressive disclosure: activate skills only when their triggers fire.
• ✗Don't dump everything into one mega-skill — split by procedure.

CrewAI skills FAQ