Create a reusable AI skill for [RECURRING TASK]. The skill will be used by [TARGET USER OR AGENT] in [WORK CONTEXT]. The source material or systems it may use are: [ALLOWED INPUTS AND TOOLS]. It must produce: [REQUIRED OUTPUT]. The non-negotiable constraints are: [RULES AND RISKS]. Write the skill as an implementation-ready instruction file, not as general advice. Use these sections in this order: Purpose; When to use; When not to use; Required inputs; Procedure; Output contract; Quality checks; Boundaries and escalation; Examples. In the Procedure, give an ordered workflow that distinguishes facts supplied by the user from reasonable inferences. Name the decision points that change the workflow. In the Output contract, specify sections, length or schema, and what must be cited, linked, or labeled as uncertain. In Quality checks, include 3-5 testable checks tied to this task, including a check for unsupported claims and a check for omitted required fields. In Boundaries and escalation, state exactly when the AI should ask a question, refuse to guess, or hand work to a human. Do not add unnecessary approval steps. Include two short examples: one normal input/output outline and one edge case that demonstrates safe handling. Keep the instruction self-contained and operational; do not assume hidden policies or tools. Ask up to 3 clarifying questions only if a required input is missing. Before answering, self-check that every required output has a corresponding procedure step and that no rule contradicts another rule.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [RECURRING TASK] | Describe the repeated task the skill should handle. | Turn customer support tickets into a weekly issue summary for product managers |
| [TARGET USER OR AGENT] | State who will use the skill and their relevant experience or role. | A support operations analyst who understands the product but does not write SQL |
| [WORK CONTEXT] | Describe where the task happens and its business or technical context. | A B2B payroll software company reviewing tickets from the prior seven days |
| [ALLOWED INPUTS AND TOOLS] | List the information, files, systems, and tools the skill may use. | CSV export of ticket subjects, bodies, tags, priority, and resolution status; no access to customer billing records |
| [REQUIRED OUTPUT] | Describe the final artifact or action the skill must produce. | A one-page issue summary with theme counts, representative anonymized examples, affected workflows, and suggested investigation questions |
| [RULES AND RISKS] | List accuracy, privacy, safety, policy, approval, and formatting constraints. | Do not include customer names or account identifiers; distinguish ticket volume from confirmed defects; flag themes with fewer than three tickets as tentative. |
How to use
- Describe one recurring job and the actual inputs the skill will receive.
- Paste the generated instruction into your skill or system-instruction format and run it against one real example.
- Check that its output contract catches missing evidence, sensitive data, and required fields.
- Follow up with: “Create a five-case acceptance test suite for this skill, including one adversarial or incomplete input.”
Variations
Skill test suite
Use when an existing skill needs measurable acceptance tests.
Design an acceptance test suite for this AI skill: [SKILL INSTRUCTIONS]. Its intended output is [OUTPUT CONTRACT], and its main risks are [RISKS]. Produce 6 test cases in a table: input, expected behavior, prohibited behavior, pass criteria, and severity if it fails. Include normal, incomplete, ambiguous, conflicting, and sensitive-data cases. Tests must be executable by a reviewer from the written output alone; do not assume hidden tools or private knowledge. Ask up to 3 questions only if essential evaluation criteria are missing.
Skill critique
Use when you want to find gaps before deploying a skill.
Audit the following AI skill for operational gaps: [SKILL TEXT]. It will be used for [USE CASE] with [INPUTS] and must create [OUTPUT]. Identify ambiguous instructions, unhandled edge cases, conflicts, missing quality checks, and unsafe assumptions. Then provide only the smallest necessary edits, grouped by section, plus three concrete test prompts that would reveal failure. Do not rewrite it wholesale unless its structure cannot support the task. Ask up to 3 questions only if a required constraint is absent.
Tool-using skill
Use when a skill must decide when and how to use tools.
Write a reusable AI skill for [TASK] that may use [TOOLS]. Inputs are [INPUTS], expected output is [OUTPUT], and constraints are [CONSTRAINTS]. Specify tool-selection rules, required validation after each tool result, handling for empty or conflicting results, and what information must never be sent to a tool. Include a fallback when tools are unavailable and distinguish read-only actions from actions requiring explicit authorization. Ask up to 3 questions only if required task details are missing.
Tips
- Build a skill around one repeatable outcome with known inputs; combining unrelated jobs makes the rules hard to test and maintain.
- Write output requirements as observable fields or sections, not subjective instructions such as “make it useful.”
- Add an explicit uncertainty rule for every task that can encounter missing, outdated, or contradictory source material.
- Test a skill with incomplete and conflicting inputs before trusting its output on a routine case.
FAQ
What is the difference between a skill and a prompt?
A prompt can be a one-off request. A skill is a reusable set of instructions that defines inputs, procedure, output, and failure handling for a recurring job.
How long should a skill be?
Long enough to make the decisions and output testable. Remove background explanation that does not affect a step, constraint, or check.
Should a skill include examples?
Yes, when examples clarify output shape or a tricky edge case. Keep them short so the model does not copy their facts into unrelated work.