Constraints beat adjectives
In every run log on this site you'll see the same pattern: prompts with concrete constraints score 10–15 points higher than adjective prompts. 'Sleek' is unmeasurable; '16px radius, near-black #07080B canvas, one violet→cyan accent used only for CTAs' is a spec.
- Give the palette as hex/HSL, and say where each colour is *allowed* — restraint instructions land harder than colour names.
- Give a component inventory (nav, bento features, pricing toggle) so the model doesn't invent sections.
- Give a motion budget ('entry reveals + one ambient pulse, nothing else') so it doesn't animate everything.
- Say what NOT to do ('no autoplay video, no stock handshakes') — prohibitions are the most reliable instruction class.
- Write like you know the subject. A real-estate prompt that names 'sqm, transit minutes, brass accents' beats 'modern apartment' by a mile.
The skeleton that scores 90+
prompt skeleton
Build a [page type] in [stack].
Palette
• canvas / ink / accent (hex) — and where each is allowed
• accent reserved for [moments]
Above the fold
• [section]: [2-3 concrete requirements, incl. typography scale]
• [headline copy — write it for me, in this voice: ___]
Sections (in order)
1. [name] — [what it must contain + interaction]
2. [name] — ...
3. [name] — ...
Motion budget
• entrances: [easing/duration hint]
• interactive: [hover/click rules]
• prohibited: [list]
Tone & traps
• write copy like [persona]
• avoid [stock clichés]
• accessibility: [WCAG notes]Why we test on three models
A prompt is a contract with a model — and models differ. In our runs, GLM-4.6 often nails editorial restraint, Codex nails technical demos, Claude balances both but loves decorative gradients unless told not to. That's why every prompt here ships with a run log: you see which model to run it on before you waste a session.
Steal our format
Copy any prompt from the library and adapt the Palette/Motion-budget sections to your own design system. The structure transfers; the content is yours.