inhabited-design is an open Claude Code skill. It designs a landing page by simulating the page's actual skeptical reader, critiques its own work in their voice, and revises until that in-character reader scores it 70/70. Below is a real run: one model's first pass on the left, the converged page on the right. Poke either one.
no signup · no demo video · the receipts are inline · install is an open Claude Code plugin
use the {skill} skill, let's create a landing page for an energy drink targeting finance bros
only {skill} changes per run — impeccable · frontend-design · inhabited-design. The Lovable control got the brief without the skill line.
Dark #0a0a0a hero, green glow, italic-serif accent, gradient chart. Competent — and indistinguishable from a thousand vibe-coded pages. No reader was ever consulted.
Same brief. Matte-black card, a machined run-gauge, a serial struck into the metal, one rust accent held to two text nodes. Built for "Tanner, 26, M&A analyst" — and rated by him, in character, until it cleared 70/70.
A designer agent proposes. A second agent — the simulated target reader — leaves inline critique. The designer revises. The thread resolves when the reader's score stops moving. This is the actual CARRY run, iteration 2 → 3, transcribed from the saved feedback files.
"It's a run of clean, correct, in-group moments rather than one holy-sh*t beat… at rest, on first scroll, no single frame makes me audibly react. It rewards me for touching it; it doesn't ambush me."
Added an at-rest beat — a Solari split-flap DEPARTURES board that flips on load and on scroll-in, resolving to the CARRY itinerary. Reduced-motion-safe: it renders at final text instantly.
"Crossed the line. The split-flap is the beat… the frame I'd screenshot into the chat before I even tap anything. At-rest ambush plus correct-to-the-cent checkout. I'm buying a case."
resolved · +3 points · thread closed at 70/70source: id_vs/exp3/iter_2/runner_feedback.md → iter_3/runner_feedback.md · verbatim, not paraphrased
The score isn't a vanity number; it's the resolved state of seven rubric checks the in-character reader ran. Here is the VANTA run's final scorecard, each check beside the receipt that closed it.
The glyph carries the meaning, not the color — [✓] reads as passing with the hue stripped. Each 10 is anchored to its own line in the verdict; e.g. Originality cleared on:
"A caffeine can sold as a numbered allotment instrument, with a machined bezel gauge and a serial struck into the metal. I have not seen this. It does not smell like a dark-SaaS template."
— id_vs/exp2/final/runner_feedback.md
Everything above audits the BPS / VANTA / CARRY runs. But this page was built the same way, for "Devin" — a skeptical engineer — and held to the same gates. So here is this artifact's own scorecard, sealed where it's resolved and openly in-progress where it isn't. The page does not get to hide the part that's still being reviewed.
— v1_critic/critic_verdict_2.md · the seam edge means: this gate is resolved
This is the honest part. The kentō notch (not a seam) marks a join that isn't closed: an in-character "Devin" is still reviewing the very page you're reading, and the last few points are still moving. The day it reads 70/70, the notch becomes a seam — until then it stays a notch.
— the loop you are currently inside · proof-state shown beside final (kyōgō)
inhabited-design-vs draws each input — framing, designer voice, seeds, competitors, type — from an enumerated pool with a millisecond-clock RNG seed. Reproduce any pick yourself; the page is its own audit log.
This very page's designer voice — Anni Albers, weaving-as-notational-grid — was drawn, not chosen. Pool of 20, K=1.
random.Random(1780861420574).sample(range(20),1)
The one-shot's tells, removed line by line. The hero rebuild as a real unified diff — glyph-coded, not color-coded.
-1,9 +1,8 @@ .hero .hero { background: radial-gradient(#6d28d9,#0a0a0a); color: #a3a3a3; background: var(--paper); color: var(--ink); /* 13.4:1 */ text-shadow: 0 0 40px #8b5cf6; .proof-pair__seam { background:var(--seam) }
| run / brief | critic cycles | final |
|---|---|---|
| BPS · energy drink | 2 | 70/70 |
| VANTA · energy drink | 3 | 70/70 |
| CARRY · energy drink | 2 | 70/70 |
| all three, identical brief | 7 | 210/210 |
| check | iter 2 | iter 3 | Δ |
|---|---|---|---|
| Originality | 9 | 10 | +1 |
| Visual Impressiveness | 9 | 10 | +1 |
| Wow Factor | 9 | 10 | +1 |
| total | 66 | 70 | +4 |
figures are real, drawn from each run's saved runner_feedback.md and critic_verdict_*.md. The Δ column carries a sign, not a color.
The whole pipeline, animated. The receipts above are single frames of a longer machine. This walkthrough runs it end to end on VANTA's saved audit trail — seven sampled inputs → a seven-image inspiration bank → the critic and ICP loops closing to 70/70. Every pool, millisecond seed, ms-sample command, score and screenshot is read straight from the run; hover any candidate for its reason, click any pass to enlarge and compare it.
step through the VANTA pipeline →Each thumbnail below is the live final/index.html from a real run — not a screenshot of one. Each is sealed as a verified 70/70 run. Click through and poke it; the source is one interaction away.
The contrast is process, not scoreboard. The one-shot pages above were generated in a single pass with no reader and no loop, so there is no score to compare against — that would be dishonest. What changed is the method: a simulated reader was in the room, and the page was revised until that reader stopped objecting.
one-shot: 1 model pass, 0 reviews, 0 scores · inhabited-design: designer↔reader loop, 2–3 critic cycles, converged at 70/70 on every run
The three above were looped: a simulated reader sat in the room and the page was revised until they stopped objecting, then it was sealed. The seven below fired once — one model pass, no reader-in-the-loop, no rubric — so they carry no seal and no number (inventing a 70/70 for them would be the exact dishonesty this page is built against). What's worth seeing is how much they agree: handed "an energy drink for a finance bro," three different tools independently reached for the same surface tells — a scrolling stock ticker, a "STRONG BUY" rating, a glowing can, a $-prefixed wordmark, trading-desk copy. The looped three skip all of it. Open them side by side — the repeats are the point.
What you are comparing is process, not points. None of the seven above were scored, so there is no number to set against the sealed 70/70 — that comparison would be invented. What you can read off them is how much they agree: hand the same brief to three different tools and, un-looped, they reach for the same half-dozen moves — the scrolling ticker, the "STRONG BUY," the glowing can, the trading-desk copy. The sealed three skip every one. The only variable that moved the needle was whether a reader sat in the room: the sealed pages were critiqued and revised in-character until the objections stopped; these seven fired once and were never reviewed. Same brief, same caffeine, same fictional analyst buyer — the loop is the difference.
sealed three: designer↔reader loop · 2–3 cycles · converged + sealed 70/70 · these seven: 1 pass · 0 reviews · 0 score · unsealed by design
inhabited-design is an open Claude Code skill. The plugin marketplace is the short path: three lines and you're running it on your own brief. No account, no trial, no funnel.
recommended · plugin marketplace
/plugin marketplace add Shimin-Zhang/inhabited-design
/plugin install inhabited-design@inhabited-design
/reload-plugins
Then run /inhabited-design:inhabited-design — interactive by default; append auto to run unattended, or lite for a single budget pass.
or · manual install
git clone https://github.com/Shimin-Zhang/inhabited-design.git
cp -r inhabited-design/skills/inhabited-design ~/.claude/skills/
No "Start building free." No demo video. If the receipts above didn't earn it, a button wouldn't either.
Strip the cans away and the machine you just read is an eval loop: a rubric, a simulated judge in the reader's voice, and a gate that won't seal until the score stops moving. Aim it at a landing page and you get VANTA. Aim it at a shipping AI feature and you get the thing that matters more — a number that tells you whether the last change broke it.
That second one is the day job. The same loop — rubric, judge, gate, revise until the score stops moving — is what Wolf Peach Labs builds for AI product teams. A feature you can't measure is a feature you can't change.