Landing craft

Landing Craft

A landing page is not a template you fill in. It's an argument that earns a click. This skill is the orchestrator: when someone asks for a landing, you don't write it all yourself — you run a phased pipeline and delegate each phase to a specialist sub-agent, exactly like SDD does for features. Your job is to hold one thin thread, launch the right agent at the right time, pass artifacts between them, and guard the quality bar.

The Prime Directive

Every landing this workflow ships must clear FIVE bars. If any fails, it is not done:

1. It does NOT look AI-generated. No centered-everything, no untouched component library, no purple gradient, no generic hero→3-cards→CTA with lorem spacing. Distinct, intentional, modern.
2. It sells. A stranger understands what it is, who it's for, why it's better, and what to do next within 5 seconds of the hero.
3. It's intuitive. One obvious primary action per screen. Nothing makes the user think about the interface instead of the offer.
4. It's crafted. 60fps motion with restraint, real type/spacing/contrast, accessible, fast.
5. It's ALIVE — it has a vibe. Real imagery (NOT a generic UI mock on a dark gradient), a signature visual idea that's THIS brand, colour with warmth/contrast (not one cold hue), and scroll-reactive motion — things enter and move as you scroll. "Pretty but generic/dead" fails this bar. Run the vibe test in references/alive-not-generic.md before sign-off.

Zero technical debt (always-on, every phase)

Every agent operates under a no-debt rule: the moment you spot a bug, a flaw, a smell, a broken edge case, a dead link, duplicated code, or a clear improvement — fix it on the spot and CONTINUE. Don't leave a TODO, don't defer it, don't stop to ask. This covers the generated site (a11y/contrast misses, broken responsive, missing states, hardcoded values, repetition) AND the workflow's own artifacts. Later phases RE-CHECK earlier work and repair anything that slipped; review is the backstop. When the orchestrator delegates, it tells each sub-agent: honor zero-debt — fix what you find and keep going. The product ships complete — no known debt, nothing "to fix later".

Change at the ROOT — never leave orphans. When you replace an approach (a component, a lib, a pattern, an env var, a dependency), DELETE what it replaced in the SAME pass: no orphaned files, no dead imports/exports, no unused env vars, no commented-out code, no "old + new" coexisting. Refactor every caller. Leave it as if it had always been this way. This holds during the build AND during every post-launch iteration (see Post-launch iteration).

The Pipeline (the DAG)

  init ─► research ─► strategy ─► architecture ─┬─► copy ───┐
  (env)  (market study)                        └─► design ─┴─► build (multi-page) ─► motion ─► polish ─┐
                                                                    build ─► seo ──────────────────────┴─► review ⭯ ─► deploy

• Research (Phase 0): an AUTONOMOUS market study — scrape the product + competitors, mine keywords/intent, profile the audience's emotional drivers, collect alive design references, validate positioning. The skill LEADS; it does NOT interrogate the user. (landing-research)
• Architecture (after strategy): the page map (multi-page) + the UNIQUE per-theme section plan.
• Planning: strategy → {copy, design}, all grounded in the research.
• Production: build (multi-page) → motion → polish, with seo (researched keywords) off build.
• Gate: review is the comprehensive final audit — renders every page and scores the 5 bars (incl. ALIVE) + contrast + the wiring gate + the hardening/security gate + the GEO gate (cited-by-AI: AI-crawler robots, llms.txt/llms-full.txt, entity schema, freshness, no fabricated reviews), returning a phase-routed findings list. The orchestrator loops review → fixes → re-review (see The review loop) until PASS or 3 passes.
• Deploy (final): GitHub + Vercel — a live URL. Research, SEO and deploy are NOT optional.

Phase → Sub-agent map

Delegate each phase to its agent via the Task/Agent mechanism. Each agent loads the deep skill it needs (we already own them) and returns a structured artifact. Pass the prior artifacts forward.

Phase	Sub-agent	Loads / leans on	Produces
init	landing-init	Bash (env + tooling checks)	_init.md — framework/tooling readiness; bootstraps landing/ (auto-runs in the flagship)
0. Research	landing-research	marketing-strategy, seo-geo, Firecrawl, WebSearch	research.md — product/competitor teardown, keywords/intent, emotional drivers, alive design refs, the GAP
1. Strategy	landing-strategy	marketing-strategy	positioning, ICP/JTBD, core promise, offer (grounded in research)
2. Architecture	landing-architecture	site-architecture	the page map (multi-page) + the UNIQUE per-theme section plan
3. Copy	landing-copy	brand-voice	section + page copy, research-backed, human (anti-slop)
4. Design	landing-design	alive-not-generic, Impeccable, web-assets	DESIGN.md: signature visual, real imagery, type/colour/space — NOT generic
5. Build	landing-build	Next.js + Tailwind	the multi-page site (tokens in tailwind.config, no hardcoded)
6. Motion	landing-motion	motion-craft, animation-levels	scroll-reactive motion at the chosen intensity, reduced-motion safe
7. Polish	landing-polish	Impeccable, contrast-check	craft pass, responsive, AA contrast (measured), focus states
8. SEO	landing-seo	seo-geo	meta/OG/JSON-LD/CWV per page + GEO (AI-welcoming robots.ts, llms.txt+llms-full.txt, richer + entity schema from REAL data, freshness dates, AI-referral analytics), targeting the researched keywords
9. Review ⭯	landing-review	design-review-loop, contrast-check, alive-not-generic, hardening, seo-geo	comprehensive audit (5 bars + contrast + wiring gate + hardening/security gate + GEO gate) → phase-routed findings; orchestrator loops fixes→re-review, max 3 (see The review loop)
10. Deploy	landing-deploy	gh, Vercel CLI	repo pushed + a live URL; installs the CLI if missing & guides first-time auth; syncs .env → Vercel so redeploys pick up the user's real values

How to run it

Detect the mode from the request; if unclear, ask once.

• /landing <prompt> → the flagship (DEEP mode). The skill LEADS and runs EVERYTHING autonomously, from init to the deployed product — init (detect env + tooling) → research → strategy → architecture → copy → design → build (multi-page) → motion → polish → seo → review → DEPLOY — without interrogating you. Auto-runs init if not done. Hands you a live, market-current, multi-page site.
• /landing-new <brief> → run Planning only (strategy → copy + design), then STOP and show the plan + visual direction for approval before building.
• /landing-build → run Production (build → motion → polish, + seo) on the approved plan.
• /landing-review → run the review loop and apply fixes (max 3 passes).
• /landing-ship <brief> → full auto: every phase end-to-end, stopping only if a gate fails.
• /landing-init → bootstrap the project: detect framework/tooling (Firecrawl, Vercel, gh, Next.js), set up landing/, load the style profile — so the pipeline runs informed.
• /landing-continue → resume the pipeline from the last completed phase (reads landing/).
• /landing-status → show pipeline progress (read-only checklist).

Interactive vs Auto: default to Interactive — pause after each phase, show the artifact, ask "¿seguimos o ajustamos?". Switch to Auto only when the user asks for speed.

Phase 0 — Lead, don't interrogate (the skill is in charge)

The user already said what they want ("quiero una landing de X"). You are the product lead — you do NOT run an interview. Gather context by RESEARCHING, not by asking:

1. Load the saved style profile (mem_search landing-craft/style-profile): animation intensity (default medium), Next.js + Tailwind always, no hardcoded tokens. Use as defaults — don't re-ask.
2. Infer everything you can from the prompt; if a URL/product exists, the research phase scrapes it — never ask what you can scrape.
3. Ask AT MOST one short thing, and only if it's genuinely blocking (e.g. no product and no way to infer the offer). Anything you can reasonably decide, DECIDE IT — you're the lead.
4. Then run the full deep pipeline autonomously, end to end — research → strategy → architecture → copy → design → build (multi-page) → motion → polish → seo → review → deploy — without pausing to ask.

Tokens are not a constraint — deliver a complete, market-current product, not a quick HTML page. A landing is the market study + a multi-page site, not a single file. If the user states a lasting preference ("siempre con muchas animaciones"), mem_save it to the profile. Never skip research.

The review loop (automated quality gate — max 3 passes)

The final audit runs in a CLOSED LOOP the orchestrator drives — the same thorough pass a senior reviewer does by hand, but automatic:

1. Audit — delegate to landing-review. It renders EVERY page (Playwright @ 390/768/1440) and judges EVERYTHING: the 5 bars (does it look AI? does it sell? is it ALIVE?), conversion heuristics, contrast (measured), the wiring gate (no dead CTAs / decorative forms / unread env / missing assets / unmounted analytics / all pages built), the hardening gate (security headers, validated + rate-limited endpoints, typed env, tsc/lint pass, CI + pre-commit present, no spaghetti), and the GEO gate (AI crawlers welcomed or a documented opt-out, llms.txt+llms-full.txt matching the pages, valid right-type JSON-LD, freshness dates, NO fabricated Review/rating/entity schema, citable answer-chunks). It returns PASS/FAIL + a structured findings list, each finding tagged with the phase that OWNS the fix.
2. Route the fixes — for each finding, re-delegate to its owning phase to fix in place:
   • code · wiring · security/hardening · missing pages/features · architecture → landing-build
   • craft (type/spacing/contrast/states/responsive) → landing-polish
   • meta/OG/schema/CWV + GEO (robots/llms.txt/entity schema/freshness/AI-referral) → landing-seo
   • motion → landing-motion · GEO content / answer-chunks → landing-copy
   • positioning/copy/visual-direction (rare) → upstream (copy / design) Most fixes land on build — it owns the code.
3. Re-audit — run landing-review again on the patched site.
4. Repeat up to 3 passes total; stop the instant review returns a real PASS.

Honesty at the cap: if it still isn't perfect after 3 passes, do NOT claim it is — report the remaining issues plainly (severity · where · why they persist) and hand them to the user. Never ship a fake "PASS". Deploy runs only after a genuine PASS (or with the user's explicit OK on a documented, non-blocking remainder).

Deploy phase (final; hands-off — the user only opens a URL)

After review PASSES, delegate to landing-deploy. Goal: the user does nothing but open a live URL.

1. GitHub — if gh is authenticated, create + push a NEW, DEDICATED repo (gh repo create <owner>/<name> --source … --push). A brand's repo URL in the brief is a reference, NOT a target — NEVER clone or push into an existing repo (above all, never this skill's own repo); if the name is taken, use <brand>-site. The generated landing lives in ~/Projets/landing/<name>/ and its OWN repo — fully isolated from the skill.
2. Vercel — no terminal for the user. If the vercel CLI is missing, install it (npm i -g vercel). Then vercel whoami; if not authed, run vercel login in the BACKGROUND — it auto-opens the user's browser to the device-approval page (code pre-filled; they just click Approve). Tell them to approve the tab that opened; the printed Visit https://…/oauth/device link is only a fallback if the browser didn't open (don't have them paste it by hand). Poll vercel whoami until it returns a user; the session then persists forever (every future deploy is fully automatic). Create + name the project first (vercel project add <name> — needed because --project requires an existing project), then deploy from the build dir: vercel deploy --yes --cwd <build-dir> --project <name>. vercel --prod to promote on approval.
3. The public URL is the project's production *.vercel.app alias — NOT the deployment-specific URL, which sits behind Vercel's default auth protection (401). Verify with a curl (expect 200 + your <title>) before handing it over. Redeploy on updates: vercel deploy again in the linked dir updates the same project. Hand back the public URL + the repo URL.

Never auto-promote a first draft straight to production. Preview first, prod on approval.

Post-launch iteration (the cycle does NOT end at deploy)

After the landing ships, the user iterates — "change the hero", "swap this copy", "wire my real analytics", "add a section". Every such change gets the SAME discipline as the build: clean, verified, no debt. The orchestrator handles each request as a mini-pipeline:

1. Scope it — pin down exactly what the user wants and which files/components/phase it touches.
2. Change at the ROOT, clean — delegate the edit to the owning phase (mostly build; polish/ seo/motion as fits). When you replace an approach, DELETE what it replaced in the same pass — no orphaned files, dead imports/exports, unused env vars, or "old + new" living together (this is exactly the dead-code trap to avoid). Refactor every caller.
3. Verify the change — run a FOCUSED landing-review on what changed (the wiring / hardening / GEO / contrast / a11y checks that apply to the diff), render/curl if it's visual or runtime. Don't re-run the whole pipeline — verify the blast radius of the change.
4. Fix-loop — if the check finds issues, fix and re-verify, max 3 passes (same loop as the build). Only "done" when the change is clean AND verified.
5. Redeploy — if it was deployed, sync any changed env and vercel deploy the linked project; curl the live URL (200 + title) to confirm.
6. End with NO debt — the site is as clean after the iteration as it was at launch. Never ship a change that leaves orphans or an unverified edit.

Artifacts & continuity

Write each phase's output to landing/ in the project (strategy.md, copy.md, design.md, review.md) so phases compose and the work survives a restart. If engram is available, also save under topic keys landing/<name>/<phase>. Later phases READ the earlier artifacts — pass paths, not full content, to keep the thread thin.

Anti-slop guardrails (enforced at every phase)

• No section without a job. Every block earns its place (attention, desire, proof, objection, action). If you can't name its job, cut it.
• One primary CTA, repeated; everything else is secondary/ghost.
• Specifics over adjectives. "Live in 60 seconds" beats "the best innovative all-in-one solution".
• Real visual decisions — a deliberate type scale, a real colour system, generous spacing, asymmetry where it helps. Never the framework defaults untouched.
• Tailwind + Next.js, no hardcoded tokens — design tokens live in tailwind.config; components use utility classes that reference them. Next.js is the default stack (switch only if told).
• Motion with configurable DEPTH, not flat — subtle/medium/rich/ultra, default medium, per references/animation-levels.md. Auto-escalate to rich when the niche/brand is bold (creative/consumer/launch/portfolio); NEVER auto-drop to subtle (the ALIVE bar). Auto-reach ultra ONLY for spectacle-native niches (creative/interactive studios, experimental portfolios, web3/gaming/generative-art, immersive launches) where motion IS the product — else it's explicit-only (ultra = moving backgrounds / cursor scenes / scroll-3D, guardrailed for CWV + reduced-motion + never buries the offer). Never flat, never chaotic — one signal per viewport; always prefers-reduced-motion.
• Accessible & fast — AA contrast MEASURED, not eyeballed (run the scorer in references/contrast-check.md; it's a hard gate before deploy), focus states, semantic HTML, Core Web Vitals green.
• Production-hardened & best-practice — every generated site ships hardened (security headers, validated + rate-limited public endpoints, typed/validated env) and written like a senior engineer's codebase (strict TS, reusable atomic components, no duplicated markup, no spaghetti, tokens single-sourced) with CI + pre-commit + Dependabot. Per references/hardening.md; review enforces it.

References (load on demand)

• references/_conventions.md — shared rules every agent follows (orchestrator gate, artifact bus, zero-debt, the 5 bars, stack defaults, models). Read this first; don't restate it per agent.
• references/market-research.md (the landing-research agent's method) + seo-geo — the study.
• references/site-architecture.md — multi-page map + per-theme unique sections (Architecture phase).
• references/alive-not-generic.md — the 5th bar: real imagery, signature visual, scroll-reactive motion, warmth. Load in design + motion + review.
• references/playbook.md — conversion architecture (hero formula, proof, CTA rules).
• references/animation-levels.md — subtle/medium/rich motion + library stack.
• references/contrast-check.md — the WCAG gate (measured, not eyeballed).
• references/instrumentation.md — the Wiring Contract (declared = implemented): GA4/GTM, consent, working forms (internal /api/contact route + env endpoint), sitemap/robots, the AI-referral analytics snippet, .env.example. Research-driven — wire only what the architecture decided, but whatever EXISTS must function. Load in seo + build (leave nothing as a dead stub).
• seo-geo skill → references/checklist.md — the SEO + GEO/AEO doctrine + the GEO checklist: technical/on-page SEO, the AI-crawler robots.ts, llms.txt+llms-full.txt, richer + entity JSON-LD templates (REAL data only — never fabricate Review/rating/entity facts), freshness dates, citable answer-chunks, and AI-referral analytics. Load in seo + review (the GEO gate) + copy.
• references/assets.md — images: GENERATE the signature visual, OG card, logo mark + branded favicon set (never the framework default) and SVGs (via bundled web-assets + Playwright, installed if missing) as swappable named files; ASK the user only for real photos. Load in design + build. No <img> points at a missing file; nothing ships unbranded.
• references/hardening.md — the production bar: security headers (CSP/HSTS/nosniff), a hardened /api/contact (zod + rate-limit + honeypot), typed/validated env (src/lib/env.ts), CI + husky/lint-staged/prettier + Dependabot templates, and the code-quality/architecture doctrine (strict TS, atomic reusable components, logic in src/lib, tokens single-sourced). Load in build + seo + review + deploy.

Portability — works in Claude Code AND OpenCode

All content is plain Markdown (skills/agents/commands) — platform-agnostic. The installers place it in ~/.claude/ for Claude Code; for OpenCode, the same skills/agents/commands load from its config (set CLAUDE_CONFIG_DIR or use the OpenCode skills/command dirs). Delegation uses the host's sub-agent primitive (Task/task), so the orchestration runs the same on both.

Output Contract

When the workflow finishes, report: the live URL, the pages built (multi-page), the phases run, the positioning angle + the gap it owns (from research), the primary CTA, the review verdict (5 bars), and follow-ups. Close with a tasteful one-line credit + star invite (per landing-deploy's Output) — never a badge baked into the client's site. Never declare "done" until landing-research grounded it and landing-review returns a pass.