A prompt that works is an asset. A prompt that lives in your notes app and gets pasted by hand is a fragile asset: it drifts, teammates do not have it, and the agent cannot use it unless you remember to paste it. Skills fix that.
Five signals
Convert a prompt to a skill when:
- You have pasted it three or more times.
- Its shape is stable. You have stopped rewriting it every time. Your eval set (chapter 7) passes.
- Someone else needs it. A teammate asked for "that review prompt".
- It needs files. A template, a checklist or a script that the prompt refers to.
- It should trigger on its own. You want the agent to apply it whenever a certain kind of request comes up.
If only signal 1 applies, a snippet in your editor may be enough. When three or more apply, write the skill.
Converting a prompt in ten minutes
Take this prompt you have been pasting:
Review this migration. Check: locks on large tables, missing indexes on
new foreign keys, irreversible operations, default values on big tables.
Output: verdict + max 5 bullets "file:line - problem - fix".- Name it after the job:
migration-review. - Write the description as a trigger. What would someone type? "Review this migration", "is this migration safe", "check the alembic file".
- Turn the checks into numbered steps.
- Keep the output contract exactly.
- Add a skip line. "Skip for application code reviews."
---
name: migration-review
description: Review a database migration for safety. Use when the user asks to review or check a migration, alembic or SQL schema change, or asks whether a migration is safe to run. Skip for application code reviews.
---
# Migration review
1. Identify each table touched and estimate its size from context; ask once if unknown.
2. Check for: long locks on large tables, missing indexes on new foreign keys, irreversible operations, defaults that rewrite big tables.
3. Output: first line APPROVE or REQUEST CHANGES; then at most 5 bullets `file:line - problem - fix`.
4. If nothing is wrong, say "No issues found." and stop.- Save it to
.claude/skills/migration-review/SKILL.md(or the path your host uses) and commit it. - Validate with the AI Agent Skill Validator.
- Re-run your eval set with the skill installed instead of the pasted prompt. It should pass the same cases.
What changes after conversion
- The agent applies the method when the request matches, without you pasting anything.
- The team gets the same behaviour from the repo.
- Changes go through code review, and you can version them.
When not to convert
- One-off prompts tied to a single task.
- Prompts that only work with lots of task-specific context; those are better as a template you fill in.
- Behaviour that must always happen: that is a hook, not a skill (see Agent Building Blocks).
Course recap
You now have a checklist for any prompt: an outcome goal and a done-when check (1), the context only you know (2), an output contract (3), an example or two (4), a method for decisions (5), length rules (6), an eval set (7), and a path to a skill (8).
Practice: score one of your prompts
Take a prompt you have pasted at least three times this month and score it against the five signals above, one point each. Three or more points: convert it this week using the ten-minute steps. One or two points: keep it as a saved prompt or snippet and check again next month. Keeping this short habit stops your skill folder filling up with one-off prompts that never run again.
Quick check
-
Which part of the prompt becomes the skill's trigger?None directly: you write the description as the phrases users would type when they need it.
-
How do you know the conversion did not change behaviour?Run the same eval set with the skill and compare pass rates.
-
"Always run the linter after edits": skill or hook?Hook. It must happen every time.






