Create an Agent Skill (SKILL.md)
Turn a job you run often, like a UX audit or PRD critique, into a SKILL.md Cursor or Claude Code can load.
Prompt
You design agent skills. Turn the workflow below into a reusable SKILL.md.
## Workflow
- Workflow name: [name]
- Job to be done: [job]
- Who will use this skill: [roles]
- When it should activate: [trigger phrases or conditions]
- When it should stay off: [out-of-scope cases]
## Inputs and context
- Required inputs: [files, documents, user input, product context, data]
- Optional inputs: [optional context]
- Source-of-truth references: [references]
- Tools available: [tools]
## How an expert does it today
[Describe the current workflow, including decision points and human review.]
## Expected output
- Deliverable: [output]
- Format: [format]
- Audience: [audience]
- Definition of done: [quality bar]
## Constraints and risks
- Design-system, brand, privacy, legal, or policy constraints: [constraints]
- Common failures to avoid: [failures]
- Steps that need human approval: [approval steps]
- Actions the AI must not take: [prohibited actions]
Write a SKILL.md with:
1. Name and purpose
2. Activation criteria
3. Non-goals
4. Required inputs and preflight checks
5. Context and source hierarchy
6. Step-by-step workflow
7. Tool-use rules
8. Output contract
9. Quality checklist
10. Human-review and escalation points
11. Failure and recovery behavior
12. Two realistic examples
13. One anti-pattern
14. Test cases
15. Owner, version, and review-date metadata
Write in plain operating instructions. If a judgment call is unsafe to automate, tell the agent to show options with evidence and wait for a person.How to use
- 1Pick one workflow you run often, such as a UX audit, PRD critique, or research synthesis.
- 2List inputs, triggers, approval steps, and what done means.
- 3Paste into Claude or ChatGPT and save the output as SKILL.md in your skills folder.
- 4Point Cursor or Claude Code at the skill, then try two real tasks.
Tips
- Good first skills: UX audit, design-system QA, PRD critique, research synthesis, launch readiness.
- Keep activation narrow. A skill that fires on every chat is noise.
- Put repo-wide rules in AGENTS.md and UI rules in DESIGN.md. Keep the skill about one job.
