Writing a brief the model can't ignore

The anatomy of a brief that reproduces

A brief that reproduces has four layers, and each answers a question the model cannot avoid. Context: what is this page for, and who is it for? Constraints: what is forbidden (no gradients, no stock copy, English only)? Structure: what blocks exist, in what order, with what content obligations? Fidelity markers: what would a great version include that a lazy version would skip?

Motif's format encodes all four in every prompt, and the run logs bear it out: briefs with an explicit block list reproduce structure at 90%+ fidelity; briefs that describe 'a modern, clean landing page' reproduce a genre, not a design.

• The context line names the industry and the audience in one breath: 'a bike shop page where the service menu reads like a menu a mechanic would stand behind' — the metaphor does the work a paragraph of adjectives cannot.
• The block list is a contract: '• Service tiers: three named tiers as cards, each listing what is actually done' — the model now has a checklist it can be scored against, and scoring is what makes the next iteration possible.
• The constraint list is negative space: 'never from $X with no date', 'no stock sunset photography' — models fill constraints with surprising creativity; they fill vague praise with clichés.
• The palette line is a decision pre-made: 'workshop grey, safety orange accent, honest grease on the cards' — colour direction is where vague briefs collapse into generic gradients.

The three sentences that matter most

In every brief there are three sentences the model treats as load-bearing. The first line (what the page is) sets the genre. The first bullet of the first block (the hero's job) sets the hierarchy. And the last line (the palette or the restraint rule) sets the taste ceiling. Rewrite those three and you can keep the other 90% identical while changing the output completely — which is exactly what the retry loop exploits.