A Practical Prompt Iteration Workflow

Overview

Unstructured prompt iteration loops forever: you change three things at once, output improves, and you don't know which change did it — so the next edit breaks it again. The fix is the same one engineering applied to code: small revisions, one intent per change, reviewed before adoption. A diff enforces the discipline — if the diff shows five unrelated changes, the revision is too big to learn from. The loaded pair shows one well-shaped iteration: a single intent (add output control) executed cleanly.

How to use this resource

1. 1

   Declare one intent

   This iteration's intent: add output-shape control. Nothing else changes. The loaded pair executes exactly that.

2. 2

   Diff before adopting

   The report should show only changes matching the intent — here, two additions, both format control, zero removals.

3. 3

   Adopt or reject as a unit

   Clean diff plus better output = adopt v2 and it becomes the new baseline. Messy diff = split the revision and redo.

4. 4

   Keep the chain

   Each kept report documents one iteration. Three reports later you know precisely how v1 became v3 — and why.

Why This Works

• One intent per revision is what makes cause and effect learnable in prompt iteration
• The diff is an honesty check: it shows what you actually changed, not what you meant to change
• Kept reports turn iteration history into reusable knowledge instead of folklore

Use cases

• Improving a working prompt without risking what already works
• Learning which single change actually moved output quality
• Running a disciplined v1 → v2 → v3 improvement loop

FAQ