Where everyone gets stuck
Generate → not happy → "make the headline bigger" → "change the colours" → "remove that section" → twenty follow-ups later you have a page nobody can explain, and the constraints you set at the start are long gone.
Follow-ups are not free. Every one adds to the context, and adherence to early instructions decays as the context grows. The type scale you fixed in turn 3 has quietly drifted by turn 15.
The test: which layer is wrong
Sort the problem into three layers. They call for completely different responses.
| Layer | Example | Response |
|---|---|---|
| Structure | Wrong section order, a whole block missing, wrong page type | Fix the prompt, regenerate |
| Style | Inconsistent palette, muddled type hierarchy, uneven spacing | Fix the prompt, regenerate |
| Local | One card's copy is weak, one button sits wrong | Follow up, scoped |
The rule is short: structure and style go back into the prompt; only local problems deserve a follow-up.
The reason is that structure and style are global properties. Ask for "change the accent to green" and the model updates the places it remembers, leaving the rest inconsistent — and that inconsistency is harder to fix than the original problem.
Why regenerating feels wrong
It feels like throwing away work.
It isn't. The real output of round one was never the page — it was the discovery of what your prompt was missing. Fold that back in and round two starts far ahead of round one.
A practical rhythm:
- Write the prompt → generate
- Don't follow up yet. Just list every problem
- Translate every structural and stylistic problem into new prompt constraints
- Regenerate
- Now only local issues remain → start following up
Most pages settle by round two. Everything after that is trimming.
How to write a good follow-up
Once you're in follow-up territory, there's still a right and wrong way.
Bad — stacked modifiers in one sentence:
make the hero nicer, bigger headline, punchier button, and fix the spacing
Good — one thing, specific values, and restate the constraints:
Change the hero only: headline from 40px to 48px, keep line height 1.1. Leave all other sections untouched. Keep the palette defined earlier.
"Leave all other sections untouched" is worth writing every single time. Collateral edits are the main way follow-up loops go off the rails.
When to abandon the thread entirely
Any of these means stop patching:
- You're past 8–10 follow-ups
- The model fixes one thing and breaks another
- It's forgotten your technical constraints (an external UI library reappeared)
- You can no longer explain how the current version relates to the prompt
Describe the parts you like, combine them with your original constraints into a fresh prompt, and open a new conversation. Much faster than continuing to patch.
What makes one-shot possible
One-shot success isn't luck — it's sufficient information. Back to the six-part structure: product, audience, sections, style, colour, constraints. Prompts with all six land first time far more often than prompts written on feel.
Follow-ups are patches. The prompt is the source.
Prompts on MotionSites are the finished result of exactly this loop — the iteration is what you're skipping.