Iteration: Nobody Prompts Right the First Time
Armed with four pages of craft, Tessa writes her most careful prompt yet — task, audience, shape, two examples, three constraints — and the draft that comes back is... seventy percent right. Good bones, wrong emphasis, one paragraph that misses the point. Her instinct is that she failed: a better prompter would have nailed it. That instinct is wrong, and unlearning it is this page. Nobody prompts right the first time — not beginners, not professionals, not the people who build these systems. The craft is not the perfect first ask. The craft is what you do with the seventy percent.
Reframe the draft as what it actually is: raw material that arrived in six seconds. Working with a model is like working with a fast, tireless junior copywriter — the first version lands almost instantly, and your editing notes are what turn it into the finished piece. Professionals expect two or three rounds and get there quickly. Perfect-first-try hunters write ever-longer prompts, get frustrated, and take longer. The iterators win, consistently, and this page is their method.
Steer with Specifics
The quality of a revision round is set by the quality of the steering, and steering has the same rule as everything in this chapter: specific beats vague. "Make it better" is barely a steer at all — better how, by whose lights? "Shorter" is weak. Strong steering points at parts by name and says what to do with them: "Cut the second paragraph entirely. Keep the pricing line exactly as is. The opening should mention the lake in the first sentence." The model revises surgically when told where to cut.
And say what was right, not just what was wrong. "The tone is exactly right — keep it" protects the good parts through the next revision. Without it, a revision can fix your complaint and lose the thing you loved, because nothing marked it as loved.
Continue the Chat, or Edit the Prompt?
Two different moves hide inside "try again," and choosing between them is a genuine skill. Continue the conversation when you are refining: the draft is close, and the context of what came before helps the next round land. This is the default, and it is usually right. Go back and edit the original prompt when the approach itself was wrong — wrong angle, wrong framing, a misunderstanding baked in from the start. Continuing from a failed approach drags the failure along in the window (Chapter 1: everything in view influences what comes next); a corrected prompt starts the assembly clean, without the wreckage.
The tell: if you are fixing details, continue. If you are fixing the idea of the answer, edit and rerun. Tessa's missed-the-point paragraph is a detail — she continues. Had the whole draft sold the wrong tour, she would edit the prompt that let it.
Know When You Are Done
Iteration has diminishing returns, and the professionals feel for the curve. Round one fixes the big things; round two polishes; round three is often nitpicking that a human edit would finish faster. When two rounds in a row stop improving matters, the remaining gap is usually yours to close by hand — and closing it by hand is not a defeat. The model produced the ninety percent in a minute; the last ten percent is where your judgement belongs anyway.
That closes the chapter's arc, and it is worth seeing whole. Say the task precisely. Supply what only you know. Show the pattern. Show the shape. Then steer, in rounds, and know when to stop. Five pages, one discipline — and one honest gap remaining: all of this makes answers fit better. None of it makes answers true. A beautifully steered draft can still recommend a hotel that does not exist. What to do about that is the next chapter, and it is the most important one in the book.
- "A good prompter nails it in one." Nobody does — professionals expect rounds and are simply better at diagnosing what to change next. The 70% draft is the process working, not failing.
- "If the first answer is off, the model can't do this task." It often can, one steer away. Abandon after informed rounds have failed, not after one vague attempt.
- "Restarting fresh is always cleaner." Restarting throws away context that was helping. Reserve the prompt-edit for when the approach failed; continue the chat when only details need fixing.
- "Iterating until perfect is the professional standard." The professional standard is feeling the curve: when rounds stop improving matters, finish by hand — the last ten percent is where your judgement belongs anyway.
- Iteration is the habit that makes the rest of the chapter practical — you rarely deploy every technique perfectly up front, and you never need to.
- The continue-or-edit choice and the stopping point are what separate productive use from an hour of thrash — the difference is not talent, it is knowing these two decisions exist.
Knowledge Check
What does this page say a 70%-right first draft means?
- The prompt was weak — a skilled prompter would have got it right first try
- The process is working — raw material arrived fast, and steering finishes it
- The task sits near the model's limits, so the whole thing would be faster by hand
- The randomness dial was set too high for this kind of careful task
Which steering message is strongest, by this page's standard?
- "This isn't quite right — make it better and more professional"
- "Make the whole thing shorter and tighter, and keep it professional"
- "Cut paragraph two. Keep the pricing line as is. Open with the lake in sentence one"
- "Try again, and this time really focus on doing your best work"
When should you edit the original prompt and rerun, instead of continuing the chat?
- Whenever any detail in the draft needs changing, however small the fix looks
- When the approach itself was wrong, so the failure should not stay in view
- After every second round of steering, to stop the chat from filling the window
- When the model starts to sound tired of being corrected in the same long chat
Two rounds in a row have stopped improving the draft. What is the professional move?
- Keep iterating — persistence is what separates professionals from amateurs
- Start a completely fresh chat and rebuild the request from zero with what you learned
- Discard the draft — if iteration has stalled, the output is unusable
- Finish the remaining gap by hand — that last stretch is where your judgement belongs
You got correct