The first response is a draft, not a verdict on the model. Result is the closing loop: compare what came back with the goal you stated, name the specific gap, and iterate on that gap — rather than rewriting the prompt from scratch and losing everything that already worked.
Check
- Against your own criteria. You stated the goal and format in the prompt; grade the answer on exactly those, not on general impression.
- Verify the facts. Claims, citations, numbers and APIs need checking against the source — fluency is not accuracy.
- Run it. For code, SQL or configuration, execution is the fastest and least forgiving reviewer.
Refine
- Change one thing at a time. Add the missing constraint and re-run. Rewriting everything makes it impossible to tell what helped.
- Correct with specifics. "Too long" is weak feedback; "cut section 2 to three sentences and drop the examples" is actionable.
- Keep what worked. Quote the good paragraph back and ask for the rest to match it.
- Promote a winning prompt. When a phrasing consistently works, save it as a template — a reusable prompt is an asset.
Know when to restart
- Wrong premise. If the model misread the task at the root, further patching compounds the error — restate the task cleanly.
- Polluted context. After many corrections the thread may contain contradictions; start fresh with the refined prompt.
Example
Vague
That's not what I wanted, try again.
Better
Good: the structure and the tone in section 1 — keep them. Wrong: section 2 lists features, but I asked for customer benefits. Rewrite section 2 only, as 3 benefit statements of one sentence each. Leave the rest untouched.