Higgsfield · 6 concepts · 4 models · each prompted to its own strengths · 2026-08-07
Six directions, generated with a different prompt strategy per model rather than one prompt reused six times. Green means worth building from; red means it fails on inspection even where it looks good. Read the “watch out” lines before picking — two of these are handsome and wrong, and that is the failure mode worth catching now rather than after it is built.
Prompted for this model: Creative-director prose with named influences (Linear, Sentry) and material/light language, palette left free. This model treats brand names as influence rather than instruction, so naming them is safe here.
The strongest result. Same density as the current build, but it breathes — three tonal surface steps, real padding, section labels in tiny caps, data in monospace. Confidence shown as bars rather than words. One red row pulls the eye immediately.
Watch out: It invented fields the product deliberately does not collect — date of birth, gender, pregnancy status. Take the layout and the tonal system; never the content.
Prompted for this model: Same creative-director framing, but pushed toward typographic restraint and negative space rather than density.
Argument first: a large quiet headline, one deck line, and a single slab of evidence beneath. The interrupt marker on the timeline with its hairline callout is exactly the thesis made visual — our code cut the model off, at a known moment.
Watch out: Reads as a marketing page rather than a working tool. Strong as an opening screen or an act header; it cannot carry the whole demo alone.
Prompted for this model: Same model, directed toward light surfaces, daylight, and a licensable healthcare-software register.
Genuinely credible practice software — calm, legible, urgency as a coloured dot, a plain action per row. Sells "this could ship tomorrow" better than anything else here.
Watch out: It invented a whole competing brand (CareDental, Riverside Dental Group) and a full nav the product does not have. Also trades away the engineering-console credibility the current design earns.
Prompted for this model: Deliberately zero brand names — this model takes references literally and has previously branded a mockup with whatever product the prompt cited. Pure structural description instead.
Handsome at a glance, and wrong underneath. It rendered a 911 dispatch console rather than dental intake, because "emergency triage" was read literally.
Watch out: The encoding fails: the timeline axis runs to 60 minutes for a call the transcript shows lasting 43 seconds, call #1025 appears twice, and "Ambulance Er Route" is a typo. A plausible-looking wrong chart survives a glance — which is exactly why it should not be built from.
Prompted for this model: Given the editorial brief specifically, since this model has produced credible landing pages before while failing console briefs. A genuine re-test rather than an assumption.
The serif headline is lovely and correctly spelled.
Watch out: Everything below it disintegrates: the subhead reads "AI-assisteel liercolisions" and the transcript is pure nonsense. It holds a headline and collapses on any text-dense panel — so it is unusable for a product whose whole subject is a transcript.
Prompted for this model: The only model accepting an explicit locked palette, and the only one reliably spelling supplied strings. Given every on-screen string verbatim and the real theme hexes — its actual strength, rather than asking it to invent a design.
Every specified string is exactly right: the wordmark, all three tabs, the panel header, all three tier pills, all three names and times. Zero drift. That confirms the rule — supplied text is exact, invented text drifts.
Watch out: Beautiful spelling, thin design. It buys legibility by giving up depth, and left the CAPTURED panel empty. Useful for a real asset with fixed copy; not a direction to build from.
Image generation locks a look, never a spec. Every on-screen string here is invented except in concept 06, there is no design-system consistency between renders, and the layouts do not correspond to real data. Pick a direction; I translate it into real tokens and components in code, where it becomes consistent and correct.
Cost: roughly 12 of 633 Higgsfield credits.