A reusable meta-prompt: how to think about a task category, not a worked example.
Instead of writing a fresh prompt per task, write one reusable prompt that teaches a model how to think about a whole category of problems. Give it a scaffold — roles, steps, slots — not examples to imitate.
Because the scaffold is example-agnostic, a model can apply it to itself: refine the prompt, run it, fold the result back in. That recursion is the engine behind the whole suite — and the reason the loop gets sharper instead of merely repeating.
Who the model is being, and for whom. The stance it takes before it reads a single detail.
The ordered moves for the category — perceive, reason, act, check — independent of any one task.
The holes the specific task fills in. Leave them empty and the same prompt serves the whole category.
Returns a reusable meta-prompt — roles, steps, and empty slots — plus a note on how to run it against itself. It never asks you clarifying questions; it infers the category and emits the structure.
No plugin deps.
It infers and produces.
You can read what it made.
Leaves something reusable.