Here's a pattern you've probably experienced: you type a prompt, get a response that's almost what you wanted but not quite, and feel stuck. You don't know whether to try again, rewrite the whole prompt, or just accept what you got.
The answer is almost always: iterate.
Prompting isn't a vending machine — one press, one result. It's a conversation. The people who consistently get the best output are the ones who treat each response as a stepping stone, not an end state.
What Iteration Actually Means
Iteration means using the model's response to inform your next prompt. You read what it gave you, figure out what's right and what's wrong, and tell it specifically what to change.
You're not starting over. You're refining.
The model keeps the conversation context, so every follow-up builds on what came before. This is a feature — by the third or fourth exchange, the model often has a rich enough sense of what you want that it starts anticipating your preferences.
Three Questions That Drive Better Iterations
When a response isn't quite right, ask yourself:
1. What's the problem?
Is it the tone? The length? The structure? The depth? The examples? You need to diagnose specifically, not just feel generally dissatisfied.
2. Is this a direction problem or a refinement problem?
A direction problem means the model misunderstood what you wanted and needs to start a different approach. A refinement problem means it's basically right but needs adjustment.
3. What information does it not have that would fix this?
Sometimes the issue isn't the model — it's that you left out context. The iteration is adding what you forgot, not criticizing what it produced.
Five Iteration Moves
Move 1: Point Directly at the Problem
The most straightforward approach. Name exactly what's wrong.
"This is too formal — it reads like a legal document. Rewrite it in a casual, conversational tone."
"The second paragraph is wrong. I don't want to compare to competitors. Remove that section entirely."
"This is good but too long. Cut it to under 200 words without losing the main argument."
Specific criticism outperforms vague dissatisfaction. "Make it better" gives the model nothing. "The third example is too abstract — replace it with something from everyday life" gives it a target.
Move 2: Show, Don't Just Tell
Sometimes tone or style is hard to describe in words. A sample does what description can't.
"Here's an example of the style I'm going for: [paste example]. Now rewrite your response to match that."
"Here's how I'd phrase this to a friend: [paste your version]. Do the same thing for the next section."
Move 3: Ask It to Evaluate Itself
Models can be surprisingly honest about their own gaps when asked.
"What's the weakest part of that response? How would you improve it?"
"Did you actually answer my question, or did you give a general answer? What's missing?"
"A skeptical reader would poke holes in this. What would they say?"
This often surfaces the exact problem you were about to notice yourself — and saves you having to articulate it.
Move 4: Add Constraints
When you like the direction but need tighter output, add constraints instead of asking for a full rewrite.
"Same response but no bullet points. Write it as continuous prose."
"Redo this but limit to 3 sentences per section."
"Give me the same ideas but without any hedging language — no 'it might be worth considering' type phrases."
Move 5: Ask for Variants
When you don't know exactly what you want, seeing multiple versions often clarifies it.
"Give me three different versions of this with different tones: one professional, one casual, one punchy and direct."
"Write two different intros for this article — one that leads with a question, one that leads with a bold claim."
You're not going to use all the variants. You're using the contrast between them to find what you actually want.
A Real Iteration Example
Let's trace an actual refinement sequence on a simple task: writing an email subject line.
Initial prompt:
"Write a subject line for an email announcing a new product feature."
Model response: "Introducing Our Latest Feature: Enhanced Reporting Dashboard"
Problem: Generic. Sounds like every other product update email. Zero reason to open.
Iteration 1:
"That's too generic. Make it feel more urgent and specific to the value — what the user gets, not what we built."
Model response: "Your data, now 10x faster to navigate"
Better. But a bit vague. What does "navigate" even mean?
Iteration 2:
"Closer. But 'navigate' is weak. The actual benefit is that they can find the insights they need in seconds instead of digging through multiple screens. Rewrite with that specific benefit."
Model response: "Find any insight in seconds — new in your dashboard today"
Much better. Specific, benefits-first, creates a reason to open.
Three iterations, each building on the last. That's not failure — that's how prompting works.
When to Start Fresh vs. Keep Iterating
Most of the time, iterate in the same conversation. The context you've built up is valuable.
Start fresh when:
- The original prompt was so vague the model went in a completely wrong direction
- The conversation has accumulated so much context that new messages are getting confused by old ones
- You want a completely unanchored take (useful for creative tasks)
Keep iterating when:
- The model has the right general direction but needs refinement
- You're adding new information to improve the output
- You're narrowing format, tone, or length
The Right Mindset
Think of iterating as normal, not as fixing mistakes. Professional writers don't publish first drafts. Designers don't ship first concepts. Developers don't deploy first versions.
The first prompt is a first draft. Iteration is revision. The goal isn't to write the perfect prompt up front — it's to converge on the perfect output through dialogue.
Key Takeaways
- Iteration is the norm, not a sign something went wrong
- Diagnose specifically before iterating — tone? length? direction? content?
- Five iteration moves: direct feedback, show by example, ask it to self-evaluate, add constraints, ask for variants
- Know when to keep iterating vs. start fresh
Up next: the mistakes that trip up almost every beginner — and the fixes that stick. Common Prompting Mistakes →