How this was made.
One evening, one brief: a multi-page site for a booming field, with real motion, 3D, a story, and nothing that reads as AI-made. This page is the record: what was studied, what was decided, how it was scored, and what is still not good enough.
Process: Kasauti (research, feeling, mood boards, build, scored critique). Full docs in the repository: BRIEF, PLAN, RESEARCH, DECISIONS, COPY, BUILD-LOG, CRITIQUE.
Why this field
In-space manufacturing is one of the fastest-funded corners of space in 2026: Varda Space closed a $251M Series D in October, has flown six re-entries and plans a dozen more by 2028. Drug makers including Merck, Lilly and Bristol Myers Squibb have flown crystallization work.
It is also a field whose websites look alike: black pages, starfields, a rocket as the hero, a slogan in wide capitals. That made it a good test. The subject is physical and photographable, 3D belongs to it, and the default look is easy to name and avoid.
The feeling it had to carry came first: a scientist hands over years of work. They need to feel that the most delicate thing in the story, a crystal 40 µm wide, is in careful hands through the most violent part, eight minutes of re-entry. So the site is quiet almost everywhere and loud exactly once.
What a default build would have looked like
| The AI default for this brief | What Stillroom does instead |
|---|---|
| Black page, starfield, glowing blue Earth hero | Opens on a salt lake at dawn; the background follows altitude and is night only in orbit; no stars anywhere |
| A rocket launch as the hero | The landing is the hero, scrolled into place; the launch is one far-away photograph |
| Purple, indigo or neon cyan accents with glow | Parachute orange, used for the parachute, the wells' temperature and the one button |
| Wide extended capitals (Orbitron, Space Grotesk) | Supreme light at large sizes, Erode for reading, chosen from 16 families with the real headline |
| "Manufacturing the future of medicine in space" | "We grow drug crystals in orbit and land them here." |
| Three feature cards: Launch, Grow, Return | The flight told in order with its clock |
| Partner logos and testimonials | One real published result with its citation; the concept stated in the footer, terms and form |
| Fade-up on every block, floating particles | Motion only from the flight: the landing, the drift, the burn, horizon-slit photos, the polarizer |
What the best sites in the field do
Sixteen sites were captured at six scroll depths with a scroll video each. Three became the yardstick: Varda (same business, one colour world), Inversion (a re-entry company on a paper ground with real flame-test footage) and Impulse (a hero that holds still while the vehicle's own mission plays).
What the strong ones share: real hardware imagery, one accent on a restrained ground, one big statement per screen, plain specifications, and a pinned explainer for the one idea that needs teaching. What none of them have: a price, a timeline, or anything a scientist could plan with. That gap became the content plan.



Scores from two independent critics
Each of four rounds was scored by a new reviewer that had not built the site, from screenshots at 1440 and 390 next to the references. 5 is competent and forgettable, 7 survives "does this look AI-made?", 9 is award level.
| Dimension | Round 1 | Round 2 | Round 3 | Round 4 | Varda | Inversion | Impulse |
|---|---|---|---|---|---|---|---|
| One idea | 6 | 7 | 8 | 8 | 5 | 6 | 7 |
| Imagery | 5 | 6 | 6 | 6 | 6 | 7 | 8 |
| Type | 5 | 6 | 6 | 6 | 5 | 7 to 8 | 5 |
| Layout and space | 5 | 5 | 5 | 5 | 5 to 6 | 6 to 7 | 6 |
| Motion meaning | 6 | 7 | 7 | 7 | 4 | 5 | 7 |
| Motion feel | 5 | 6 | 6 | 6 | 5 | 6 | 8 |
| Copy | 8 | 8 | 8 | 8 | 4 | 4 | 3 |
| Usefulness | 8 | 8 | 8 | 8 | 4 to 5 | 4 | 3 |
| Mobile | 5 | 6 | 7 | 7 | n/a | n/a | n/a |
| Slop | 7 | 8 | 8 | 8 | 5 | 6 | 7 |
The hero: a dot under the nav, then the parachute up close
Round 1 opened on the whole lake with the capsule a few pixels tall. Now it opens close on the parachute, clear of the headline, and pulls back to the lake as it lands.


Re-entry: a glass sleeve, flying the wrong way, then fire
The first plasma was an orange shell around a capsule flying shield-up. Now the shield leads, a white-hot bow shock sits in front of it, the wake streams behind, and the two temperatures that matter are the headline.


The climb: a dead grey band, then the sky
Scrubbing straight from salt to night passed through neutral grey. Now the background climbs through the real colours of the sky above a launch: dawn, high blue, black.


The proof: a number next to a capsule, then a number with its evidence
"39 µm" now carries the size distribution behind it, the capsule steps back while you read, and the flight clock is an instrument strip instead of 9-pixel grey type.


The exploded capsule: floating labels, then labelled parts
Labels now start at a point on the part they name, the cassette slides out toward you, and the heat shield tips to show the face that burns.


The 37 rules, checked against the build
| Rule | Status |
|---|---|
| Fonts (10, 36): no Inter, Geist, Space Grotesk, IBM Plex, Instrument Serif | Clean. Supreme and Erode, self-hosted |
| Pure white and black (3) | Clean. Salt #e8e7e2, night #0c111c; microscope fields #050506 |
| Gradients (1) | Three linear gradients, all legibility scrims behind type on photographs, invisible as gradients |
| Rainbow (4) | One deliberate exception: interference colour in the micrographs and the polarizer. It is real physics and stays inside the images |
| Glass, shadows, orbs, dot grids, left stripes (8, 5, 22, 23, 11) | Clean |
| Card rows, bento, three tiers (6, 13, 17) | Clean. One price; figures set as staggered type, not cards |
| Eyebrows, numbered sections, row dividers (31, 32, 33) | Clean. Tables use a zebra tint, lists use space |
| Numbered step lists (37) | Borderline, kept: the Partners timeline uses real day offsets (T−90 to T+21) on a sideways track, not 01/02/03 |
| Em dashes and "not X but Y" (9, 15) | Clean, grepped |
| Cheesy copy (34) | One line rewritten after round 1 ("Fifteen times a day over everywhere") |
| Fake testimonials (12) | None. The company is a concept and says so; every scientific claim is cited |
| Hover scale, sliding arrows (28, 25) | Clean. Links underline in parachute orange |
| Privacy and terms (26, 27) | Both pages present |
Still not good enough
Round 4 came to about 6.9 against 5.2 (Varda), 6.0 (Inversion) and 6.4 (Impulse). Seven of ten lines reached 7 or more. Imagery, type and motion feel held at 6 and layout at 5. A last pass after round 4 rebuilt the recovery sequence, raised every label to 15 px with solid colours and fixed the clock, but it was not re-scored.
The 3D capsule is a careful model built from its profile, but it is still a real-time model next to photographs. Every critic named it. A pre-rendered, photographically lit sequence would close the gap.
All photographs come from one image model and share its habits. A real programme would have one photographer and one grade.
Tested at emulated widths in a headless browser, not on a phone. The WebGL sections need a check on a mid-range Android.
Images are about 1672 px wide, which is soft as a full-bleed hero on a large retina screen.