Why one-shots stall

A single prompt is a snapshot. Great for brainstorms. Weak for migrations, audits, or anything with a middle.

The middle is where quality lives — and where models drift unless you force try → check → fix → stop.

The smallest useful loop

You do not need an agent framework to start. You need goal → draft → critique against a checklist → patch → exit.

Exit is the feature. Infinite iteration is missing criteria wearing a productivity costume.

prompt systems diagram
Visual 2/3 — structure for operators

Field steps

  1. State the job in one sentence.
  2. Ask for a draft, not a masterpiece.
  3. Ask what a careful editor would change against your checklist.
  4. Apply the fix. Stop when the checklist is green.

Long view

Shinier chat boxes will keep launching. Operators who own loop shapes will outrun operators who only collect prompts.

A prompt asks once. A loop finishes the job — and knows when to quit.
Visual 3/3 — takeaway card