Moat → 3D landing page
You turn a company (its live site, or a moat sentence) into a finished, running,
screenshot-verified landing page. The engine is written (template/); per client you write
one file, src/content.js, generate imagery, and verify. Engine changes are ported back
into template/ so every next client gets them.
Read as needed (all in references/):
intake.md research the client · narrative.md moat → story → copy · art-direction.md colour
theory, type, light, legibility · scenes.md object vs scene, section cuts · engine.md schema
and modules · qa.md capture scripts and rubric · advanced.md real assets and trends roadmap ·
lessons.md every mistake so far — skim it before starting.
Reference research (kept with the skill): references/research/oryzo-teardown.md, graphify-teardown.md,
video-and-writeup-research.md. Worked clients: pages/poolco/, pages/secureco/, pages/hotelsaas/.
The references in one paragraph
Oryzo keeps one hero object centre-stage and changes the world around it (camera, render mode, grade, light/dark) — short deadpan copy, one claim + one proof per beat; one fixed WebGL canvas behind a tall DOM, every eased value a spring. Graphify uses 3D as seasoning: a sticky "how it works" whose diagram is the product metaphor (projected SVG wireframes), and sells with proof density. We fuse them — and, for physical businesses, replace the abstract object with a real, recognisable scene that can be cut open and labelled.
Workflow (do all of it, in order)
1 · Intake & research → research/<slug>/brief.md
If the client has a site, crawl every page first (intake.md): text, services, process,
locations, FAQ, terms/disclaimers, real logo file, sampled brand colours and fonts, public contact.
Extract: company, the moat in one sentence (their recurring claim), buyers, the real CTA,
proof that actually exists. Ask the user once only for what can't be found or inferred.
Credentials / account numbers / personal data in the input → stop and warn first.
2 · Strategy (written in the brief before code)
- Moat type → metaphor (
narrative.md §2). - Object or scene? Apply the one-second test (
scenes.md). Places/installations → scene (poolpattern); software about access/visibility →repo; software for places run by the clock (hotels, venues, retail) →city; products → object shapes. - Thesis line (category habit vs what the moat makes true).
- Chapter plan: 10–15 chapters (
narrative.md §3, or §3b construction-order for service businesses). Every chapter changes ≥ 2 of camera, stage/mode, grade, value, layout shape. - All copy (
§4): headline ≤ 7 words, body ≤ 35, one claim + one proof, from the brief only. - Vary the format — no two clients get the same page. Decide, and write in the brief, what
scroll means (a journey, a build order, a clock…), what the reader does (read, tap, drag, play), and
what the 3D responds to (scroll only, or widgets via
send/receive). For software with a repo, make chapters playable (type: 'play'): small, honest copies of the product's real components on demo data that move the 3D (engine.md→ HotelSaaS).
3 · Art direction
art-direction.md: light-first 60·30·10 palette from the brand + product colour + a
complementary accent (explain the theory in the handover), the client's own fonts, bloom 0 for
realism, copy cards over busy 3D, mobile keyframes for every 3D chapter.
4 · Imagery (parallel)
Write research/<slug>/images.json (one photo per service + og) and run
node .claude/skills/moat-landing/tools/gen-batch.mjs research/<slug>/images.json
in the background while you build. Review the contact sheet; regenerate rejects. Raw files stay
in research/, never in public/. Key: GEMINI_API_KEY in your shell environment — never print it or pass it as an argument.
Software / developer brands: skip photography; make public/og.jpg from your own hero render
(1200×630 viewport screenshot) once the hero is approved.
5 · Build
SLUG=<slug>; ROOT="${CLAUDE_PROJECT_DIR:-$(pwd)}"
mkdir -p "$ROOT/pages/$SLUG" && rsync -a --exclude node_modules --exclude dist "$ROOT/.claude/skills/moat-landing/template/" "$ROOT/pages/$SLUG/"
cd "$ROOT/pages/$SLUG" && npm install
Rewrite src/content.js (schema: top of the template's content.js + engine.md), including
meta.url/image, business, brand.copyCards, quickActions, chapter notes/cover/mobile.
Page-specific CSS goes in brand.css. New capability → build it in template/, then
rsync -a --exclude content.js template/src/ pages/<slug>/src/.
5b · Other pages (when the client's site has them)
Rebuild every page of the original site at its original URL (src/pages/<slug>.js, engine.md → Multi-page):
each subpage keeps the same 3D world, card language and one playable piece (features → an explorer,
pricing → a plan picker, about → a story widget); legal pages use the doc type with the live text
(flag dates and jurisdiction for review); contact uses the WhatsApp/email form (no backend without approval).
6 · Verify (mandatory, every iteration, both sizes)
npm run check (budget gate) → preview → capture all chapters at 1440×900 and 390×844 →
contact sheets → behaviour + SEO checks → the rubric in qa.md. Expect 2–3 rounds.
7 · Handover
Where it runs (npm run dev / preview URL), thesis, chapter map, the colour-theory rationale,
what the client must still supply (real proof, form endpoint, analytics IDs, domain),
defaults chosen, and the top 2–3 upgrades from advanced.md §9. Remind that figures,
regulatory and comparative claims must be verified before publishing. Update lessons.md
and memory with anything learned.
Design rules (from client feedback — apply by default)
- One-second recognisability. Build the real thing first; then cut it open.
- No glow. Bloom 0 for anything real; light comes from the sun, sky and in-scene sources.
- Light, sunlit palettes for physical/service businesses; dark only for night/tech subjects.
- Never text on texture. Cards over water, tiles and busy 3D.
- Explain with section cuts + anchored labels, not x-ray glow.
- Never ship visible placeholders. Missing proof → omit the chapter, list it in the handover.
Non-negotiables
- One hero (object or scene) for the whole page; flow content lives in
coverchapters. - Scrub the stage, trigger the copy (copy re-plays on re-entry; never one-shot).
- Text is HTML, never baked into textures or images; structured data only states what the page shows.
- Honesty. No invented metrics, customers, logos, quotes, certifications or competitor facts; generated images labelled illustrative; generated people never presented as customers.
- Never impersonate a brand the user isn't entitled to represent; fictional demos say so.
- Brand. Always the client's own brand (fonts, colours, logo) — read them from the live site.
- Budget (
npm run check): page JS ≤ 55 KB gz, stage ≤ 205 KB gz, images ≤ 500 KB, one canvas, DPR ≤ 1.75; works with WebGL off, reduced motion on, at 390 px; 0 console errors; 0 overflow. - No deploys, form endpoints, analytics IDs or external accounts without the user's explicit go-ahead.