What is an agent skill?
An agent skill is a reusable package of instructions and supporting resources for a defined task. A skill can tell an agent when to use a workflow, which steps to follow, and which scripts or references to consult. The exact format and loading behavior depend on the agent platform, so portability should be checked rather than assumed.
What belongs in a useful skill?
A focused skill names its trigger, expected inputs, ordered procedure, output, and verification steps. References and scripts belong beside the instructions when they reduce ambiguity or repeated work.
When is a skill better than a prompt?
Use a skill when the same multi-step task recurs and benefits from stable checks or bundled resources. A one-off request with little procedure may be clearer as a direct prompt.
Skill, prompt, or tool?
| Artifact | Best suited to | Main question |
|---|---|---|
| Skill | Repeatable procedure and guidance | How should the agent do this work? |
| Prompt | A specific instruction or outcome | What should happen now? |
| Tool | A callable operation or data access | What action can the agent perform? |
Skill review checklist
- The trigger is narrow enough to avoid accidental use.
- Inputs and expected outputs are explicit.
- Risky actions include approval or verification steps.
- Platform-specific assumptions are labeled.
Frequently asked questions
Is an agent skill a plugin?
Not necessarily. A skill may be instruction-only, while a plugin can bundle broader capabilities. Terminology differs across platforms.
Does a skill execute code?
A skill can reference scripts or tools, but execution depends on the agent environment and its permissions.
Can one skill call another?
Some systems allow composed workflows, but behavior is platform-specific. Keep dependencies explicit to make the flow auditable.
How large should a skill be?
Large enough to complete one coherent job. Split it when triggers, inputs, or verification paths become meaningfully different.
How do I test a skill?
Run representative requests, edge cases, and a case where the skill should not trigger. Check both the output and the verification steps.