Fidelity 84 is a specific sentence
A low fidelity score is not 'the model failed' — it is a diagnosis with a location. The Motif rubric scores structure (blocks present, in order), content (copy is specific, not generic), and craft (details that make it feel designed). An 84 usually fails one of those three, and the fix is different for each. Structure failure means the brief's block list was ambiguous. Content failure means the copy obligations were too loose — the model filled space with filler. Craft failure means the constraint layer was missing — no palette line, no restraint rule, no 'never' list.
Run 1 fid 84 blocks present, but hero copy is generic
("Elevate your brand" energy)
Fix tightened the copy obligation: name the product,
the user, and the one verb of the hero.
Run 2 fid 91 hero specific; now the footer is filler
Fix added footer obligation + a 'no stock phrases' list.
Run 3 fid 93 shipping.The five fixes that recur
- Add a 'never' list: models default to the most generic version of a brief; 'never show a stock sunset', 'no fake testimonials', 'no 3-colour-overload' costs three lines and removes an entire failure class.
- Name the content: a block that says 'pricing cards' produces lorem prices; 'three tiers with names, real figures, and what each includes' produces a pricing section a designer could defend.
- Give the palette a job: 'calm paper, ink, one trustworthy blue' is a taste ceiling; without it the model picks its own default gradient.
- Specify the first screen: the hero is where fidelity dies most often — describe what is in it, in order, including what is NOT in it.
- Change one thing per retry: change three things between run 1 and run 2 and you will not know which one fixed it — the loop only compounds when each run is a single-variable experiment.
When to stop retrying
The loop has a natural end: when the remaining gap is craft judgment (spacing rhythm, type scale, a specific composition), a third retry costs more than ten minutes of hand-tuning the generated page. Motif's own guidance: two prompt retries max, then the generated page becomes a component you refine by hand — the retry loop improves briefs, and hand-tuning improves pages, and confusing the two is how people spend an afternoon on run 7.