Skills vs. resources
A skill is something an agent can do; a resource is something an agent can know. What each is for, the value they add, and when to reach for which.
Two building blocks are easy to confuse because both extend an agent — but they solve opposite problems. A skill is something the agent can do; a resource is something the agent can know.
At a glance
| Skill | Resource | |
|---|---|---|
| Answers | "What can the agent do?" | "What should the agent treat as true?" |
| Holds | A reusable capability / action | Ground-truth content for a turn |
| Use when | The agent must take an action or run packaged behavior | The agent must ground its answer in specific, trusted content |
| Think of it as | A tool in the agent's hands | A reference on the agent's desk |
What a skill is
A skill is a reusable capability you attach to an agent — packaged behavior the agent can invoke on its own, separate from any single prompt. Skills keep capability out of the instructions: instead of describing how to do something in a long prompt, you give the agent a skill that does it, and reuse that skill across many agents.
Reach for a skill when the agent needs to perform something repeatable. When the capability calls an external system or API, that is usually an MCP tool — see Connect an MCP tool.
Value: capability is written once and shared, agents stay small, and behavior is consistent everywhere the skill is used.
What a resource is
A resource is trusted content you hand the model as ground truth for a turn — the facts, policies, or reference material an answer should be based on. Resources reduce drift and hallucination without stuffing everything into one giant instruction block, and they can be versioned and shared across agents.
Reach for a resource when the agent needs to know something specific and stay faithful to it. Resources pair naturally with prompts — see Prompts and resources.
Value: answers stay grounded in content you control, and you can update the facts without touching the agent's instructions.
Which one do I need?
- The agent should answer from specific, trusted content → add a resource.
- The agent should take an action or run packaged behavior → add a skill (or an MCP tool for external systems).
- You're shaping tone and rules of behavior → that's neither; that's a prompt (Create a prompt).
Tip: If you're unsure, ask whether the gap is knowledge or capability. Missing knowledge → resource. Missing capability → skill.