You have a prompt in a notes file. You paste it most weeks. That is the signal that it should become a skill — but not every prompt deserves the upgrade, and doing it badly makes things worse.
The promotion test
A prompt earns becoming a skill when three things are true: you use it more than monthly, the output is wrong in ways you have to correct by hand, and the corrections are consistent. Consistent corrections are the real signal — they mean the missing piece is instruction, not intelligence.
If the corrections vary each time, a skill will not help. That is a task where the model needs better context, which means a tool, not a longer instruction.
The conversion, step by step
**Start from the prompt that already works.** Take it from the library or your notes. Do not rewrite it from scratch; you are promoting something that has evidence behind it.
**Move the variables into placeholders.** A prompt has `<diff>` and `<schema>`; a skill describes where those come from in your project. "The diff comes from `git diff --staged`" is more useful than another instruction about quality.
**Add the trigger and the refusal.** When does this skill apply, and when should the agent not use it? Without the refusal, the skill will be applied to tasks it was never designed for.
**Add the verification step.** The prompt library's entries all assume a human checks the result. A skill is supposed to run without one, so it needs its own check.
What changes when it becomes a skill
A prompt you paste is one user's habit. A skill is a claim about how the team works. That difference has consequences: someone will disagree with step three, and that disagreement needs a place to live — usually a comment in the file and a note in the repo.
Expect the first version to be wrong in a specific way: too general. Skills written by one person tend to describe principles. The fix is to append examples of the actual corrections you made, which is the most useful content a skill can contain.
When a skill is the wrong tool
If the task needs live data, a **tool** is the answer. If the task needs a decision the model cannot make, a **human approval point** is the answer. If the task is one-off, the prompt in your notes is fine — not everything needs to be infrastructure.
Back to the curriculum, or browse the skill directory to see how other people structured theirs.