ChatGPT rebuilt the page instead of polishing it
The user named a visual destination but never protected the structure and copy that were meant to survive.
What the user wanted
The actual target
Keep the existing landing page, but make its hierarchy, spacing, typography, and finish feel more premium.
What they asked
The instruction
Make this landing page look more premium.
What the AI produced
The observable result
A new layout with a large gradient hero, rewritten copy, two removed sections, and a different information hierarchy.
Why
Why WWW assigns this verdict
“Premium” describes a direction, not the boundaries of the edit. The request did not say that the layout, sections, copy, and behavior were preservation constraints, so a redesign was a plausible interpretation.
What gave it away
The changed elements were never named as protected. The model did too much, but the original instruction gave it permission to decide what “premium” meant.
Fix
What to do next
- Separate what may change from what must stay exactly as it is.
- Name the dimensions of improvement: hierarchy, spacing, typography, and visual consistency.
- Reject the visual patterns you already know you do not want.
Repaired instruction
A safer retry
Improve the existing landing page’s perceived quality while preserving its current layout, section order, copy, interactions, and responsive behavior. Change only spacing, typography, hierarchy, borders, and subtle surface treatment. Avoid gradients, new sections, rewritten copy, and generic SaaS styling. Return a concise list of every element you changed.
Model recommendation
Stay, retry, or switch?
Stay with ChatGPT for the retry. This was primarily a boundary problem, not evidence that the model category was wrong.
General lesson
What survives this one case
When the job is an edit, preservation constraints are part of the instruction—not an optional clarification after the first result.