Iterate from the output
- prompting
- iteration
You rarely get the right answer on the first try — that is normal, not failure. The skill is the second message — a prompt in its own right. Most people either restart from zero or type “make it better”; both throw away information you just paid for. The first output tells you what the model understood, and that is exactly what your follow-up should correct.
Critique, don’t redo
The draft is a map of the model’s interpretation. Use it: “Keep the two opening paragraphs. Rewrite only the closing.” Now the parts that worked are protected, and the round of changes is small enough to verify.
Four moves for a strong follow-up
- Point at the exact place. “The second bullet”, “the title” — not “some parts feel off”.
- Name the fix, not the feeling. “Cut it to three bullets” beats “make it shorter”. A feeling describes your reaction; a fix describes the target.
- Protect what works, explicitly. “Change nothing else.” Otherwise every round quietly reshuffles the pieces you liked.
- One change per round — or clearly separate ones. Five blended changes produce a rewrite where you cannot tell what helped.
When you’re stuck
Ask for options instead of an edit: “Give me three different openings for this section.” Comparing alternatives is easier than describing the fix from scratch — and you can blend the best of each in the next round. You can also ask it to list its weakest sentences as a suggestion generator; treat that list as ideas, not a verdict — the model is no better at judging its own output than at producing it.
When to restart instead
Restart when the goal changed, not the text. If you now want a different deliverable, a new prompt built from the outcome is cleaner than five patches on the wrong target.
Why it works
Every follow-up adds constraints the model can act on, and the conversation — with the draft and your corrections — becomes part of its context window, exactly as the foundations lesson explains. “Make it better” adds no constraint at all, so the model guesses; a pointed fix narrows the search to the sentence you actually mean.
One last habit for the loop: keep a two-line log of what you asked and what changed. Three rounds later the log shows which instruction actually fixed the draft — and the next similar task starts from a recipe instead of from zero. Iteration feels messy only until the record makes it repeatable.
Practice
Choose the stronger follow-up
A product announcement draft came back well structured but too long and a little salesy. You liked the opening paragraph and the three feature bullets; everything after them needs work.
Pick the follow-up message you would actually send.
Hint
Which follow-up tells the model what NOT to touch?
Count the concrete instructions each message contains.
Why this is the answer
A follow-up works when it narrows the search: keep this, cut that, fix those sentences. 'Make it better' re-rolls the dice instead of steering the draft — you get a new version to review from scratch, with no guarantee the parts you liked survive.
Transfer
- Documents: point at a section, name the edit, freeze the rest.
- Code: “keep the structure, change this function to return a list” beats “fix it”.
- Plans: “merge steps 2 and 3, keep the rest” — same loop at every scale.
Next
Next: Evaluate before you ship — the last muscle: a fast rubric you run before anything leaves your hands.