the loop · five steps
Five steps before a single pixel.
Point your AI at SaturnAI once, and every build from then on runs the same loop — study real work, commit to one direction, build with dated technique, grade the rendered page. Here is each step, in plain words.
01 · point
Point — one line. That's the whole setup.
Paste one line into whatever AI you use — no install, no account, no card. From then on, every build starts with the brain.
Everything else is a detail. The hosted brain needs no install — the line is the setup.
02 · study
Study — read what already worked.
Before it writes anything, it reads the real products in the library — what worked, what to copy, what to avoid. Named moves from studied work, not the same defaults every model reaches for on its own.
The cards name real products and real techniques, read closely enough to say what to steal and what to avoid.
03 · commit
Commit — one direction, with hard rules.
One invented look for your project, with rules attached: contrast floors, palette limits, a motion budget. One direction, binding. The build ships against it, not against a menu of moods.
One direction with hard rules is a constraint you can grade against. A menu of moods lets a build drift.
04 · build
Build — dated technique, one per slot.
View transitions, scroll-driven CSS, perceptual color, a $0 edge back end: one technique per slot, chosen from cards with a dated support line and a fallback. Not the decoration every scaffold ships by default.
Every technique card carries a dated support line and a fallback, because a technique without a fallback is a bet.
05 · grade
Grade — a verdict on the pixels.
A script your agent runs measures the rendered page — contrast math, layout logic, the known tells — and the hosted review grades that evidence: PASS, REVISE, or REJECT, with the numbers and the exact fix list. The review reads what the agent collects; it does not inspect screenshots and does not operate the page itself, so a PASS covers the evidence supplied, not every interaction. The agent that built it does not grade it.
The review returns PASS, REVISE, or REJECT, with the measured evidence: contrast values, the named AI-slop tells found in the page, and the fix for each.
06 · what a card contains
What a card contains
A card names a real product or a real technique, shows what to steal, and shows what to avoid. Techniques carry a dated support line and a fallback, because a technique without a fallback is a bet.
07 · one direction, not a menu
Why one direction beats a menu
A menu of moods lets a build drift. One named direction with contrast floors, palette limits, and a motion budget is a constraint you can grade against.
08 · the verdict, honestly
The verdict, honestly
The review reads the evidence your agent collects. The hosted review does not inspect screenshots and does not operate the page itself, so a PASS covers the evidence supplied, not every interaction. Writer is not the reviewer: the agent that built it does not grade it.
09 · what the brain refuses
What the brain refuses
Priced-in decoration with nothing behind it, contrast that fails math, motion with no budget, the tells every scaffold ships. Refusing is the product.