The problem is not the model
Two people use the same AI builder. One ships the first generation. The other iterates ten times and still gets that unmistakable template smell. The difference is almost always the same thing: the prompt was missing information.
"Build me a gym website" leaves the model to fill in 90% of the decisions from its training average — the most common palette, the safest layout, the emptiest copy. You get the average, because the average is all you asked for.
Adding information is far cheaper than regenerating. Here is the six-part structure that keeps showing up in the prompts that actually convert.
The six parts
1. Product
One sentence on what the page is. Not the industry — the page.
- ❌ A gym website
- ✅ A single-page landing page for a personal training studio, aimed at getting visitors to book a free trial session
2. Audience
Audience determines tone, density and which trust signals belong on the page.
Target audience: desk workers aged 25–40 who want to lose weight but are put off by gym sales pressure. Their main worry is being upsold into a contract.
3. Sections
This is where prompt quality is won or lost. List the blocks in order:
1. Hero: one-line value prop + booking CTA + real coaching photo
2. Three pain-point cards
3. Trial session flow (4-step timeline)
4. Coach profiles (3 people, with certifications)
5. Member results (4 items, initials only)
6. Pricing (3 tiers, middle marked "most popular")
7. FAQ (6 items, including "will I be upsold?")
8. Closing booking CTA
4. Style
Give adjectives and counter-examples:
Style: clean, restrained, generous whitespace — close to the typography of Linear's site. Do not use purple gradients, glassmorphism, or corner radii above 12px.
5. Colour and type
Never let the model pick the palette. Give values:
Primary #0F172A, accent #22C55E, background #FFFFFF / #F8FAFC. Headings Inter 700, body Inter 400 at 16px with 1.7 line height.
6. Constraints
Output format, stack, and — critically — what it must not do:
Output a Next.js App Router page with Tailwind. Use local placeholder images, never external image hosts. Do not invent numbers, awards or client names — leave
{{TBD}}placeholders instead.
That last constraint prevents the most common real-world incident in AI-built landing pages: a fabricated "trusted by 500+ companies" going live is a genuine legal exposure, not a cosmetic issue.
Put it together
Concatenate the six parts in order and you have a prompt you can paste. It will read like a spec rather than a chat message — that is correct. A landing page prompt that works usually runs 300–600 words. Anything under 150 will need rescuing through follow-ups.
When to follow up, when to rewrite
After the first generation:
| Situation | What to do |
|---|---|
| Structure is right, one block looks wrong | Follow up, scoped to that block |
| Copy tone is off | Follow up, add audience detail |
| The overall structure isn't what you wanted | Fix the prompt and regenerate |
| Palette is inconsistent | Pin hex values in the prompt, regenerate |
Ten follow-ups on a wrong skeleton is slower than one prompt rewrite — and the longer the context grows, the more reliably the model forgets your early constraints.
A template you can copy
Product: a {page type} for {business}, aimed at {conversion action}.
Audience: {who}, whose main concern is {objection}.
Sections:
1. ...
Style: {three adjectives}. Do not use {counter-examples}.
Colour & type: primary {} accent {} background {};
headings {font+weight}, body {font+size+line height}.
Constraints: output {stack}. Local placeholder images only.
Invent no numbers, awards or client names — use {{TBD}}.
Fill in the braces and the first generation will land noticeably closer.
Don't want to assemble this every time? Every prompt in the MotionSites library is already written to this structure — preview first, copy, paste.