Core Techniques

Iterating on Outputs - Critique and Refine

8 min read

Here is a mindset shift that separates frustrated AI users from productive ones: the first output is a first draft, not a verdict. You would never expect a perfect deliverable from a contractor's first pass, and the fastest AI users treat prompting the same way — as a short conversation, not a single roll of the dice.

Refine, don't restart

When an output misses, most beginners either give up or delete everything and rewrite the whole prompt. Both waste the AI's biggest advantage: it remembers the conversation (it is all in the context window), so you can steer with short follow-ups.

Too long? "Cut this to 100 words without losing the offer details."
Wrong tone? "Rewrite this warmer and less salesy — like a note to a regular customer."
Almost right? "Keep the first paragraph exactly as is. Rewrite only the ending with a stronger call to action."
Not sure what you want yet? "Give me 3 variations: one playful, one straightforward, one urgent."

Specific feedback beats general feedback here too. "Make it better" forces the model to guess what you disliked. "The second sentence sounds like a car commercial — make it plainer" fixes the actual problem.

Ask the AI to critique its own work

This is the most underused iteration move. Models are often sharper as editors than as first-drafters, and you can use that:

Here's the promo email you just wrote. Before I send it, critique
it as a skeptical customer who gets 50 marketing emails a day:

1. What would make you delete this without reading?
2. Is the offer clear within the first two lines?
3. What's the weakest sentence?

Then rewrite it fixing those problems.

Critique-then-rewrite in one prompt routinely produces a stronger result than either step alone. Dana, a freelance consultant, uses a version of this on every client proposal: draft, ask the AI to attack it as a cost-conscious client, then patch the holes it finds. Fifteen extra seconds, noticeably better proposals.

Know when to iterate and when to start fresh

Iteration is not always the answer. Use this rule of thumb:

If the output is 70% right, iterate. Point at what is wrong and refine.
If the output is fundamentally the wrong thing — wrong task, wrong audience, wrong direction — your original prompt was the problem. Start a fresh prompt and fix the Role, Context, Task, or Format (Chapter 2, Lesson 1 tells you which).
If a long chat has gone in circles, start fresh and paste in only the best version so far plus what you want changed. A clean desk beats a cluttered one.

Bank what you learn

Every iteration teaches you something about your own preferences. When a follow-up fixes a recurring annoyance — "shorter sentences," "never use exclamation marks," "always end with one question" — move that instruction into your original prompt next time. Iteration is how you discover the rules; your prompt is where the rules should end up living. This is exactly what makes the templates in Chapter 3 so effective.

Try it now

Take any AI output you were unhappy with recently — or generate a quick draft of something now. Then do three rounds: first, give one specific piece of feedback and ask for a revision. Second, ask the AI to critique its own draft from your customer's point of view and rewrite it. Third, write down the one instruction you found yourself repeating, and add it to your next prompt from the start.

Few-Shot Prompting - Show, Don't Just Tell