The gated pipeline the web-design Team Lead runs on a new marketing site or storefront — and, for a pure web-app, as a thinner brand/UX/a11y/perf wrapper around frontend-engineering's build, not the app's build pipeline: nine fail-closed gates (discovery → IA → tokens → content → build → a11y → perf → SEO/AEO → pre-launch), each against numeric WCAG 2.2 / CWV / SEO bars, seaming app builds to frontend-engineering and auth/payments to security-reviewer, ending in one Go / Conditional / No-go.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
The gated pipeline the web-design Team Lead runs on a new marketing site or storefront — and, for a pure web-app, as a thinner brand/UX/a11y/perf wrapper around frontend-engineering's build, not the app's build pipeline: nine fail-closed gates (discovery → IA → tokens → content → build → a11y → perf → SEO/AEO → pre-launch), each against numeric WCAG 2.2 / CWV / SEO bars, seaming app builds to frontend-engineering and auth/payments to security-reviewer, ending in one Go / Conditional / No-go.
Skill: gold-standard-website-pipeline
Purpose. Give the Team Lead — the orchestrator role defined in web-design/CLAUDE.md (the one role that delegates to the 7 web-design specialists; not an 8th specialist agent itself) — one ordered, fail-closed gate ladder to drive every new website build through — from greenfield discovery to launch sign-off — so nothing ships that hasn't cleared a named gate with a checkable, standards-anchored bar. It is the orchestration layer above the 12 execution skills: it says what order to run them in, who owns each gate, what the numeric pass condition is, where the pipeline parallelizes, and where it hands off — while the skills themselves say how to do the work. Used by the Team Lead (primary) and by all 7 web-design agents when reporting into a gate.
A single site build touches all 7 specialists and 12 execution skills. Left unsequenced, the failure modes are predictable: a11y audited after the build hardens (P0s cost 10× to fix), tokens invented per-component because design ran after build started, content written to fit a layout instead of the layout serving the content, "responsive" claimed from CSS inspection with no rendered check, and a launch that trails off without an explicit go/no-go. This pipeline fixes the sequence, the gate, and the seam, so each discipline lands on the inputs it actually needs and cannot pass by omission.
Standards basis (plugin-internal, durable). Every acceptance bar below traces to knowledge that ships inside this plugin, so a consumer who installs web-design carries the provenance on disk — no gate depends on a dangling link into a dated research-scratch directory:
CWV thresholds (LCP/INP/CLS @p75, field-over-lab), WCAG 2.2 AA floor + the EU Accessibility Act, image/font/third-party budgets → knowledge/web-platform-capabilities-2026.md.
Full WCAG 2.2 success-criterion set + the 5-pass audit ladder + severity guide → the accessibility-review skill; contrast arithmetic → scripts/contrast_ratio.py (runnable, CI-gateable).
Security headers (the HSTS / CSP / COOP-CORP-COEP set) + consent mechanics are encoded directly in G9 below and the plugin's templates/launch-checklist.md; consent-gating mechanics also live in third-party-script-hygiene.
Scored against the plugin's own Family A (built site) / Family B (this pipeline artifact) yardstick — both are enumerated inline in §9 (the consolidated Family-A A1–A8 dimension table and the Family-B self-check), and each Family A bar is also restated within its owning gate's acceptance criteria, so the whole yardstick travels on-disk to any consumer. Last verified: 2026-07-06 — the volatile figures (CWV thresholds, WCAG 2.2 SC, Lighthouse bands) ship continuously; re-confirm at use against the dated knowledge files above (each carries its own Last reviewed: anchor).
Derived from (one-time provenance, not a live dependency). The durable knowledge files above were reconciled during authoring against a session research pass — external standards fetched from w3.org / web.dev / developer.chrome.com / OWASP, plus a house dispatch-map/idiom review of the plugin's own agents and skills — scored against a gold-standard rubric. That derivation is a how these bars were set record, not a link any gate depends on: every gate's provenance resolves plugin-internally to the knowledge files, skills, and templates cited above, so a consumer who installs web-design carries all of it on disk. The curated exemplar set that informed the gate lessons is captured durably in-plugin at knowledge/gold-standard-website-references-2026.md.
When to use
Greenfield build — any new site or app; start at G1 before a single wireframe or line of code.
"Build me a marketing site / web app / storefront" — the Team Lead's first move on any build-shaped request. Classify the archetype at G1, then walk the gates.
Full re-platform / redesign — an existing site moving stacks, CMS, or IA runs the whole ladder (the redirect plan at G2 and the migration rows in the launch checklist exist for exactly this).
Pre-launch readiness sweep / retrofit — a build that skipped the early gates can enter at G6–G9 to run the audit-and-sign-off tail as a gate ladder onto existing work.
Periodic gold-standard re-certification of a live site (annual/quarterly) — run G6–G9 only.
Any point where "is this ready to launch?" needs an answer sharper than a verbal "looks good."
When NOT to use
A single-discipline task — "this page is slow" (→ core-web-vitals-tuning directly), "audit our contrast" (→ accessibility-review), "refactor this component" (→ frontend-implementer). Reach for the individual skill; don't spin up a 9-gate pipeline for a one-lane job.
An audit-only engagement on a site you won't rebuild — run the audit skills in parallel (a11y / CWV / content / technical-SEO) and synthesize; you don't need the discovery/IA/build gates.
App-grade product engineering with no brand/marketing surface — that is frontend-engineering's pipeline (see the web-app branch below). This skill owns brand/UX/marketing-site + the WCAG audit; it seams the app build out. Native mobile → mobile-engineering.
The whether/what of commerce (assortment, pricing, LTV:CAC, channel spend) — that is ecommerce-dtc's lane, upstream of any build. This pipeline starts when that decision has produced an actual storefront build task.
Scoping note — the pure-web-app case (read before over-scoping this pipeline onto an app). For a web app with no marketing/brand surface of its own, all nine gates still close — the thin slice is in substance, not in which gates run (G9's entry requires G3–G8 all closed or waived, so silently omitting G5/G8 would stall it). What changes: G5 closes via the frontend-engineering seam-return (they own the build; G5 passes when their build returns and still clears the G6/G7 bars), and G8 closes as an explicit N/A — authenticated views behind login, not indexable by design (marketing/landing pages fronting the app still gate at full weight). The gates carrying the real work are G1–G4 (discovery, IA, design system/tokens, in-app content/microcopy) + G6/G7 (the WCAG audit + CWV/perf verdict) + G9 (sign-off); G5 and G8 close on a seam-return and an N/A respectively — recorded, never skipped. Use this as the brand / UX / a11y / perf wrapper around someone else's build — not the app's build pipeline. If the app has no brand/UX/a11y surface you own either, this is the wrong tool: go straight to frontend-engineering.
1. Site-type adaptivity — declared once at G1, adapts every gate after
The archetype is the first output of G1 and it re-weights the ladder. Declare exactly one; a build that is genuinely two (e.g. a marketing site with a self-serve app) runs the shared gates once and forks the app/commerce concerns to the seam. Every gate runs for every site type — what changes per type is weight, N/A status, and which sub-checks apply.
Gate
Marketing site
Web app
Ecommerce / DTC
G1 Discovery
Full weight. Stack bias: static-first (Astro / 11ty / Hugo).
Full weight. Stack bias: app framework (Next / SvelteKit); state/data-layer named early. Flag the frontend-engineering handoff now if the app surface is substantial (see §6).
Full weight. Mandatory sub-step: run ecommerce-dtc's platform-vs-headless decision (themed Shopify vs. headless, 3-yr payback) before G5 Build.
G2 IA
Flat / hub-and-spoke sitemap; ≤2 clicks to any page is the common bar.
Task-flow / app-shell IA; nav models around user goals, not page hierarchy.
Catalog taxonomy + faceted nav; category/product URL taxonomy is the highest-stakes IA call (SEO + findability both ride on it).
G3 Design/Tokens
Full weight — brand is the product.
Full weight, but component-density (data tables, forms, dashboards) drives the semantic-token set more than brand expression does.
Full weight; product-card / PDP / cart component tokens are the acceptance-critical subset.
G4 Content
Full weight — editorial / landing-page copy, blog/resource density is the deliverable.
Full weight on product copy + trust-signal placement (reviews, shipping/return policy at the decision point).
G5 Build
frontend-implementer owns it end-to-end.
frontend-implementer owns the brand/marketing shell; substantial interactive surfaces escalate to frontend-engineering per the named seam — do not duplicate their build.
frontend-implementer owns a themed build; a headless/app-grade storefront escalates the storefront app to frontend-engineering, while web-design still owns PDP/cart copy + conversion-design.
G6 Accessibility
Full weight, no exception.
Full weight — a11y is never down-weighted for "MVP."
Full weight; the checkout flow is the single highest-stakes a11y surface (WCAG 3.3.8 accessible-auth applies to any login-to-buy flow).
N/A discipline (binding). When a sub-check genuinely does not apply to a site type (e.g. checkout economics on a pure marketing site, or SEO on an authenticated app view), the gate output must record N/A — <one-line reason>. A silent omission is scored as a miss, not an N/A — exactly per the Family A scoring rule (the yardstick this pipeline is scored against; Family A/B dimensions in §9).
Seam rule (house-mandated, zero exceptions). Any surface touching auth, sessions, user data, untrusted input, file upload, or payments routes through ravenclaude-core/security-reviewer regardless of archetype. Its teeth are pipeline-side, not per-agent: the mandatory, no-waiver G9 gate (and the mandatory row in the §6 seam table) that every build converges on, enforced by the Team Lead regardless of which specialist first touches the surface. Per-agent state (closed 2026-07-06):security-reviewer is now named in the escalation routes of all 7 agents — web-architect and frontend-implementer carried it before; accessibility-auditor, content-strategist, performance-engineer, ux-designer, and visual-designer each gained an archetype-appropriate security-reviewer escalation row this PR, so the routing is uniform per-agent as well as pipeline-side. The gate never depended on that per-agent routing — G9 is the enforcement, and it never waives — but the seam-rule assertion is now true at both layers. This is the one cross-cutting gate that never waives.
Rendered, browser-verified build + CI wiring; dual-analytics wired per third-party-script-hygiene §8 (empty-until-provisioned; authed/internal surfaces off by default, [TF6])
G6
Accessibility audit
accessibility-auditor (sole owner)
Each gate is fail-closed: it either passes, loops back to its owner with named findings, or takes an explicit waiver (owner + reason + condition, recorded — not a silent skip). No stage passes by omission. The ladder terminates in exactly one go/no-go verdict at G9.
3. Dependency DAG — what actually parallelizes
The ladder is not strictly linear. Two facts drive the DAG: tokens don't depend on content (G3 ∥ G4, both need only G2), and the three post-build audits are independent of each other (G6 ∥ G7 ∥ G8b, all depend only on G5). Serializing them is false serialization; a scorer credits the parallelism (yardstick B4).
Critical path: G1 → G2 → max(G3, G4) → G5 → {G6 ∥ G7 ∥ G8b} → G8 → G9. Because G5 joins on both G3 (tokens) and G4 (content), the path runs through whichever of the two parallel branches finishes later — often G3 on a marketing site ("brand is the product," full-weight design system). Shortening the schedule means shortening IA→(the slower of tokens/content)→build, not the audits (which parallelize).
Parallel branch 1 — G3 (tokens) ∥ G4 (content): both consume only the G2 IA/content-model output. A visual-designer building the token scale doesn't need final copy; a content-strategist drafting microcopy doesn't need the token scale. Running them sequentially wastes a full cycle.
Parallel branch 2 — G6 (a11y) ∥ G7 (perf) ∥ G8b (SEO technical-side): all three depend on a built surface (G5) but not on each other. They are commonly serialized by habit when nothing structurally requires it.
Join node — G5 (build): requires all of G2 (URL/template map), G3 (tokens to consume), and G4 (real copy — lorem hides layout bugs). A build that starts before these exist is running on inputs that don't exist yet — the exact failure B4 penalizes.
Split within G8 (asymmetric edge): SEO is not one monolithic gate that waits for Build. Its content-side half (G8a) — keyword/search-intent briefs, structured-data planning — opens as soon as G2's content model + G1 personas exist and is refined through G4 (only its close gates on G4's close, not its open — hence the dashed opens/closes edges in the DAG, not a hard G4 → G8a serialization); its technical-side half (G8b) — "crawlable without JS," "valid JSON-LD on the live markup" — can only close once G5 exists (those checks need real HTML). Technical-SEO is owned by web-architect, content-SEO by content-strategist; a single owner would orphan half of it (B3).
Designed-in, not bolted-on: a11y (contrast pre-checked at G3 via scripts/contrast_ratio.py, focus order baked into wireframes at G2/G4) and perf (budgets declared at G1, enforced at G7) are present from the start — G6/G7 verify what earlier gates designed in.
4. The gates in detail
Each gate states: Entry · Dispatch (agent → skill) · Acceptance criteria (checkable, pass/fail, sourced from the Family A yardstick in §9 + the plugin-internal standards basis) · Objective bar (the standards number / yardstick dimension) · Fail-closed outcome · Artifact. No gate may be marked complete by inference — every criterion is something a reviewer can point at.
G1 — Discovery & budget-setting
Entry. A request to build a new site, with no prior artifact.
Dispatch.web-architect is the lead/accountable owner of the kickoff (the stack/budget owner is answerable if the brief ships incomplete), co-owning with ux-designer and content-strategist: web-architect (stack/hosting/CDN/i18n + the numeric budgets), ux-designer (users, goals, success metrics), content-strategist (on redesign: trigger content-audit on the existing site first) jointly fill templates/design-brief.md. Each co-owns its sub-concern; the Team Lead escalates to web-architect if the brief is contested or incomplete.
Archetype declared — exactly one of marketing / web-app / ecommerce (drives §1 for every later gate + the seams).
Stack chosen with a written rationale AND ≥2 alternatives considered (not "we picked Next because Next"); static-first bias (SSG > SSR > CSR unless a reason is stated), grounded in knowledge/modern-web-stacks-2026.md. Hosting/CDN named; if Azure SWA and azure-cloud is installed, state the seam split. Web-app caveat: for the web-app archetype the hosting/CDN choice is provisional until frontend-engineering's rendering-strategy call (SSR/SSG/RSC/CSR) is made at the G5 seam — a static-first CDN picked here can be invalidated by an SSR/RSC runtime need (edge/Node). Either defer the final hosting commitment to after the G5 seam, or record it here as provisional and subject to the explicit G5→G1 loop-back defined at G5.
Perf budget declared — numeric per-template LCP/INP/CLS targets + a critical-path KB cap, written down now (enforced at G7).
— WCAG 2.2 as the floor; note any AAA stretch items adopted on primary flows.
All 5 IA artifacts present: sitemap (templates annotated), URL-taxonomy spec (slug / trailing-slash / casing / locale rules — both trailing-slash policy and case-sensitivity decided, they feed G8), navigation spec (primary/utility/footer/contextual with inclusion criteria), content model (per-template fields + relationships), redirect plan (301 chains audited for any re-architecture, old→new map before any page is retired).
Sitemap shape chosen (flat / hierarchical / hub-and-spoke / app-shell) with the one-line reason it fits the §1 site type.
Card-sort run (8–12 participants, per information-architecture §4) if this is a nav restructure rather than greenfield. Fallback when no human panel is available (e.g. a solo-founder engagement with no research budget): substitute a synthetic-persona / stakeholder-proxy sort — sort the content against the G1 audience personas and named stakeholders — and record it under the standard waiver (owner + reason + condition: no user panel — validated by proxy, re-run with real participants if budget appears). The waiver mechanism (§5) explicitly covers this criterion; a nav restructure never silently skips validation.
URL-to-template map exists (the join input G5 needs).
Heading/landmark structure implied by the IA is set up to nest without skips (feeds G6/G8 semantic checks).
Objective bar. Semantic-structure foundation (heading ranks nest, one unambiguous <h1>, landmark elements) per the heading discipline in accessibility-review + seo-technical-audit. Redirect plan mandatory on re-platform — a retired page without a 301 destroys link equity.
Coherent type / space / color / radius / shadow / motion scale — no off-scale one-offs.
Both light and dark themes defined, with an explicit dark-mode residue check (zero light-mode values surviving into dark).
Contrast pre-checked at design time with scripts/contrast_ratio.py against the actual rendered background (states flattened): body ≥ 4.5:1, large text / UI ≥ 3:1 (WCAG 1.4.3, AA floor); the AAA stretch (1.4.6, 7:1) adopted as the target on primary flows (hero, primary CTA, checkout) — so G6 confirms rather than discovers.
Components compose on shared primitives; design-system-audit runs clean — consistency + completeness, no orphan one-off styles.
G4 — Content & conversion copy (runs parallel to G3)
Entry. G2 closed.
Dispatch.content-strategist (primary) + ux-designer (conversion-copy layer) → content-audit, conversion-design. The keyword/search-intent brief refined here is part of G8a, which opened alongside G3/G4 as soon as G2's content model + G1 personas existed (see the DAG) — G8a's close gates on G4's close, its open does not.
Acceptance criteria.
content-audit run (inventory + 5-dimension scoring + KKCR matrix) if this is a redesign with existing content to triage.
Real copy for every template — no lorem, no "TBD" shipped into mocks.
Microcopy written for CTAs, errors, empty states, success messages.
Every interactive screen carries a designed ≥5-state inventory — empty / loading / error / partial / ideal — not just the happy path (yardstick A6's per-screen state-inventory bar; especially load-bearing for the web-app archetype, where loading/partial states are the ones most often skipped). A screen missing any of the five is a fail-closed loop-back, not a pass.
Trust signals placed at the decision point (reviews, guarantees, security/privacy assurances adjacent to the primary CTA / form / checkout) — archetype-agnostic per A6, not an ecommerce-only item.
Voice/tone falsifiable, not adjectival: copy conforms to the content style guide's explicit do/don't table (point at the table, not the word "consistent"), and reading level hits a stated target — Flesch-Kincaid grade ≤ the ceiling named for the G1 audience (e.g. ≤ 8 for a broad consumer site; ≤ 12 for a technical/B2B audience). Mechanized on the same tiered pattern as the rest of the pipeline: Tier 1 — run scripts/flesch_kincaid.pycheck <prose.txt> --max <ceiling> (stdlib, CI-gateable) on the plain-text copy; a non-zero exit fails the bar. — if the copy can't be extracted to plain text this pass, cap the criterion at Conditional with a named-reviewer condition (§5) to run the checker in CI. A reproducible number, never an eyeballed "reads clearly."
Entry. G3 and G4 both closed (or G4 closed with only non-critical waivers per above).
Dispatch.
Marketing (and themed storefront):frontend-implementer (primary; the only sonnet agent — the execution specialist) owns the build end-to-end. The build skill it dispatches is stack-conditional — the plugin ships two complementary named build skills, so there is a skill for each stack: fluent-react-implementation (Fluent-UI-v9 + React only by premise — its title/opening line scope it to "a Fluent UI v9 + React website end-to-end") and static-site-implementation (the generic non-React static build — semantic HTML + token-driven CSS, SSG-first Astro/11ty/Hugo/plain HTML+CSS). (The React skill's "When to use this (and when not)" Don't bullet excludes a radically-bespoke-brand project, a different constraint — do not cite it as the source of the React-only scope.)
Non-React static stack — the G1 marketing default (Astro-static / 11ty / Hugo, or any plain semantic-HTML/CSS build): → static-site-implementation (consuming G3's design-tokens-scaffolding / card-tile-ui outputs, against the same G5 acceptance criteria below); fluent-react-implementation is correctly N/A — stack is not React/Fluent. This is the explicit resolution of the former G1-recommends-static ↔ no-static-build-skill gap: the primary archetype's recommended stack now has a dedicated operationalizing skill, not just "the agent authors it directly."
Astro/Vite with React islands → fluent-react-implementation for the island components; the static shell → static-site-implementation.
All variants are verified with the cross-plugin (render → see → critique → iterate).
G6 — Accessibility audit (parallel with G7, G8b)
Entry. G5 closed.
Purpose. Every agent enforces a11y at its own layer, but the audited verdict is this agent's alone. Accessibility is P1, designed in from wireframes — G6 verifies, it isn't "phase 2."
Acceptance criteria — the checkable AA floor (WCAG 2.2).
Evidence tier — a defined ladder, so an agent with no human AT tooling still has a deterministic close path (not an infinite "conditional" loop):
Tier A — clean pass: automated (axe-core / pa11y) and manual keyboard and a human screen-reader pass (VoiceOver + NVDA min). Only Tier A can mark G6 a clean pass.
Tier B — agent-executable substitute (when the executing agent is a headless Linux/CI sandbox with no macOS VoiceOver / Windows NVDA and no human AT tester, the common fully-agentic case): automated (axe-core) + a structural AT surrogate — a Playwright accessibility-tree + ARIA computed-name/role dump and a programmatic focus-order simulation (tab-sequence + :focus-visible + focus-not-obscured checks). Tier B caps G6 at Conditional with a standing named-human-reviewer Condition (owner + target date via §5) — the same treatment as automated-only, never a silent clean pass, but a closeable state that doesn't permanently wedge G9.
Automated-only (no keyboard, no AT surrogate) → Conditional at best, weaker than Tier B.
The 6 new-at-A/AA success criteria in WCAG 2.2 — 2.4.11 focus not obscured (AA) · 2.5.7 dragging alternative / single-pointer (AA) · 2.5.8 target size ≥24×24 px (AA) · 3.3.8 accessible auth / no CAPTCHA-only, paste + password-manager supported (AA) · 3.2.6 consistent help (A) · 3.3.7 redundant entry (A) — — contrast body ≥4.5:1 / large ≥3:1 (AA, 2.1) · non-text alt (A, 2.1) · reflow to 320 px (AA, 2.1) — all verified.
G7 — Performance & Core Web Vitals (parallel with G6, G8b)
Two explicit sub-gates — because CrUX/RUM field data cannot exist for a brand-new site until it has served real traffic (the pipeline's headline "greenfield build" case). Splitting them makes the close rule unambiguous instead of leaving a structurally-unmeasurable field-data checkbox as the pre-launch bar.
G7a — pre-launch sub-gate (the mandatory close condition for a first launch; this is what G7 passes/fails on before launch):
Lighthouse ≥ 90 in Performance, Accessibility, Best Practices, SEO — the pre-launch lab proxy for the field metric, explicitly labeled as lab and never presented as equivalent. Lighthouse needs a working Chromium (via chrome-devtools-mcp or a local lighthouse/@lhci/cli), which the agent's sandbox may not have or be able to fetch — so this bar runs on the same tier ladder as G5 (a "tool unavailable" state is closeable, never a stall):
Tier 1 — live Lighthouse run (MCP lighthouse_audit, or the CLI against a local Chromium): the real ≥ 90 gate. Clean close.
Tier 2 — no reachable Chromium: substitute a static/heuristic proxy from the build's own stats output (webpack/Vite/Rollup bundle-analyzer JSON, or a dist/ byte census) — critical-path transfer size, total resource count, render-blocking-resource count, image-format/lazy-load coverage — asserted against the G1 budget. This caps G7a at Conditional with a standing named-reviewer Condition (owner: performance-engineer + target date, §5) to run the real audit once a Chromium is reachable (CI's LHCI has one). A static proxy is never a clean ≥ 90 close — it is a closeable Conditional so a Chromium-less agent never wedges.
Performance budget asserted in CI (Lighthouse-CI budget assertions), not eyeballed: ; the web.dev of critical-path < compressed on a mid-tier/3G baseline and TTI < on slow 3G applies only where G1 set no stricter number (web.dev frames 170KB/5s as starting-point defaults, not a universal mandate; the per-template budget in governs).
G8 — SEO / AEO (split ownership; G8a runs early, G8b parallel with G6/G7)
Entry. Content-side (G8a) opens alongside G3/G4 — as soon as G2's content model + G1 personas exist — and is continuously refined through G4; only its close (not its open) gates on G4's close. Technical-side (G8b) opens at G5 close. The gate closes only once both halves clear.
Semantic: heading ranks nest with no skip going deeper (h2→h4 fails), one unambiguous <h1>, landmark elements not div-soup, exactly one <main>, form inputs labeled.
Technical: valid JSON-LD describing only visible content. Pre-deploy check is scoped to what is genuinely offline-runnable — a syntactic JSON-LD parse via Bash using the Python json stdlib only (no external schema.org vocabulary): the block is valid JSON, @context resolves to schema.org, @type is present, and the required properties for that type are present per a small hand-maintained shape map (e.g. Product → name+offers; Offer → price+priceCurrency; AggregateRating → ratingValue+reviewCount). This is a presence/shape check, not full schema.org shape conformance — that is not achievable stdlib/single-dependency, so true schema.org validation is deferred to the post-deploy Rich Results Test (a JS-driven third-party web UI needing a public URL), already a standing G9 Condition. (No such validator ships in the plugin today — only — so this bar is a stdlib parse the agent writes inline at audit time, or a small shipped alongside if the plugin later adds one; it is deliberately NOT described as a schema.org-vocabulary checker it does not have.) One canonical URL per page, valid + (all URLs 200, no accidental production ), non-empty unique meta descriptions, descriptive anchor text, valid if multi-locale, and — asserted locally over the markup (the live Twitter Card Validator / Facebook Sharing Debugger are likewise post-deploy G9 confirmations, not pre-launch gates). on the (Tier 1 = live audit; Tier 2 with no reachable Chromium = assert the underlying SEO-category tags directly over the markup, capped at Conditional with a named-reviewer Condition to run the live audit in CI). // JSON-LD on PDPs.
Entry. G3–G8 all closed, or closed with a recorded waiver — a P1 may enter G9 open only with a Conditions waiver (owner + target date); a P0 never waives (§5). P2/P3 waivers are always allowed.
Dispatch. Team Lead drives templates/launch-checklist.md verbatim; each specialist signs its own block, plus the shared SEO section (G8's two owners), the Analytics / instrumentation section — accountable primary: performance-engineer (RUM/Web-Vitals wiring per G7a, page-view/conversion/error tracking, no-PII-in-analytics, and the consent-flow test that ties to §4/G9 consent mechanics), escalating tracking/tag-manager configuration to web-architect where it crosses into IA/URL/tag-deployment territory — a single accountable owner per rubric B3, not a co-owned blob, so the section is signed, not silently skippable — and two org-level sections (Legal/compliance, Communication) outside the agent roster. Mandatory, zero-exception: if the site has auth, sessions, user data, or payments (true for essentially all web-app and ecommerce archetypes, and any marketing site with a lead-capture form storing PII), ravenclaude-core/security-reviewermust sign off before Go is possible — the one cross-cutting rule with no waiver path.
Acceptance criteria.
Every specialist section signed (Signed: <name> — <YYYY-MM-DD>) — federated, not a monolithic blob: web-architect, ux-designer, visual-designer, frontend-implementer, content-strategist, accessibility-auditor, performance-engineer + the shared SEO section, the Security review section (security-reviewer — mandatory sign-off where auth/sessions/payments/PII exist, else N/A — <reason>), the Analytics / instrumentation section (accountable primary performance-engineer, signing; tracking-config escalations to web-architect co-signed only where they fired), and the org-level (Legal/compliance, Communication) sections. No section in the checklist template is left unowned — the security-reviewer sign-off is the one with no waiver path when in scope.
5. Severity, owner & target-date vocabulary (used identically at every gate)
One shared scale, reused — never reinvented per gate (yardstick B7). Every finding, at every gate (G6 a11y, G7 perf, G8 SEO alike — not accessibility alone), carries severity + named owner + target date on this scale. A bare "TODO" or unlabeled finding is not a valid gate output.
Severity
Meaning
Blocks G9 Go?
P0
Breaks a core user path, legal/compliance exposure, or a security gap.
Yes — always.
P1
Materially degrades the experience or a measurable bar (e.g. one CWV metric "Needs Improvement").
Yes — unless explicitly waived at G9 with Conditions.
P2
Noticeable but non-blocking (e.g. an AAA stretch item missed on a secondary page).
This pipeline owns brand, UX, the marketing-site / storefront-theme build, and the WCAG audit. It explicitly does not own, and hands off to:
Resolving "is that sibling plugin installed?" — operationally, never by an assumed oracle. Every "(if installed)" branch below (and the ecommerce-dtc / finance / regulatory escalations) is triggered the same way: the executing Team Lead attempts the dispatch to the target agent/skill; if it cannot be resolved — the plugin/agent isn't enabled in this session — it is treated as not-installed and the documented degraded path is taken. There is no separate "query the marketplace" step; failure-to-resolve is the not-installed signal.
Named in web-design/CLAUDE.md's "Adjacent plugins" seam as of this PR (the reciprocal ecommerce-dtc-side cross-ref is a recommended follow-up in that plugin). Treat the platform-vs-headless decision as an explicit G1 sub-step, not an assumption. Degraded path when ecommerce-dtc is not installed (a plausible split — separate marketplace plugins): web-architect drafts a lightweight platform-vs-headless memo stamped non-specialist stand-in — re-run via ecommerce-dtc once installed, and the unit-economics call is logged as an open G9 Condition (full contract in G1's escalation-flags criterion). The escalation target being absent never leaves the sub-step undefined.
Named per-agent for edge cases, routed via G4. "Installed" resolved operationally (attempt the dispatch; unresolved → not-installed); if neither is installed, route through ravenclaude-core/architect per web-design/CLAUDE.md §10.
Defer, don't duplicate. Where a web-design skill already covers a slice of a downstream concern (e.g. conversion-design's form-field-reduction guidance is directly reusable for an ecommerce checkout), this pipeline reuses it rather than re-deriving the specialist's work from scratch.
7. Evidence discipline & typed-artifact ledger (applies at every gate)
House rules a scorer checks under yardstick B6/B8 — bake them into every gate's report, not just G9.
Cited or flagged, always. Every consequential claim carries an inline citation to a source checked this session or an explicit [unverified — training knowledge] marker with a one-line reason. Volatile facts (thresholds, spec versions) carry a Last verified: <date> anchor.
Checker over restated rule. Where a rule is arithmetic (contrast, byte budget), a runnable checker stands in for prose — scripts/contrast_ratio.py for WCAG contrast, Lighthouse-CI budget assertions for perf. A gate that cannot mechanically return a fail is theater; delete it or give it teeth.
Typed, closed artifact per gate. Each gate emits the plugin's Structured Output block (---RESULT_START--- … standards_cited / budget_impact / tested_on) — machine-parseable hand-off, not free text the next gate must re-read. Each report's header states gate, dispatch owner(s), date, and a Grounding: line (a curl -I output, a Lighthouse-CI run link, an axe-core report, or an explicit [unverified] marker).
Waiver, not silent skip. A gate can be consciously waived (owner + reason + condition, recorded) — but never silently passed. The waiver is itself an artifact.
A11y audited after the build hardens. G6 as an afterthought means P0s cost 10× to fix — contrast is pre-checked at G3, focus order designed at G2/G4; G6 verifies, it isn't where accessibility originates.
Tokens invented per-component. Build (G5) starting before design/tokens (G3) → hardcoded hex scattered through the code. G3 gates G5.
Lorem in "final" mocks. Placeholder copy hides layout problems realistic text lengths would expose. G4 gates G5 on real copy.
False serialization. Running G3 after G4, or the three audits one-after-another, when they're independent — inflates the timeline and fails yardstick B4. Parallelize per the DAG.
Build on missing inputs. G5 kicked off before IA/tokens/content exist → work built on inputs that don't exist yet.
Auditing an unenforced build. Running G6/G7 against a build with no CI budget assertions — the audit passes today and silently regresses tomorrow with nothing to catch it.
Lab data treated as field truth. A green Lighthouse run declared "fast" with no CrUX/RUM p75 and no CI-enforced budget. Field beats lab; lab is only the pre-launch proxy.
SEO treated as one monolithic post-build gate. Hides the real available parallelism in its content-side half (G8a), which only needs G4, not G5.
The ecommerce platform-vs-headless decision made implicitly during G5. It must be an explicit G1 sub-step; discovering it mid-build re-litigates the stack choice under deadline pressure.
llms.txt treated as a gate. Penalizing its absence or over-crediting its presence — no vendor has committed to reading it (2026). It's insurance, not a requirement.
A P2 finding silently blocking launch indefinitely. With no waiver path, a non-critical miss behaves exactly like a P0, which trains the team to stop taking severity seriously.
Silently dropping an archetype-irrelevant sub-check instead of recording N/A — <reason> — the rubric scores a silent omission as a miss.
The audit that can't fail. A "gate" structurally incapable of returning a fail is theater — every gate here has an explicit fail path and a loop-back.
Owning everything. A pipeline with no seams that re-implements app-grade frontend, growth economics, or security review instead of handing off — duplicates other teams' work and skips the mandatory security-reviewer gate.
Trailing off without a verdict. Ending at "looks done" with no explicit . G9 is not complete without the terminal decision.
9. The two yardsticks (Family A = built site, Family B = this pipeline)
Family A — the built-site yardstick (enforced gate-by-gate in §4). The consolidated dimension table, inlined here so a consumer who installs web-design carries the whole yardstick on disk (each bar is also restated inside its owning gate's acceptance criteria):
Dim
Built-site dimension
Gold-standard bar
Owning gate(s)
A1
Accessibility conformance
WCAG 2.2 AA on core paths, no open P0/P1; AAA stretch on primary flows
G3, G6
A2
Performance & Core Web Vitals
Lighthouse ≥ 90 lab + CI byte budget + RUM (pre-launch); CWV Good @p75 field, both form factors (post-launch)
reflow 320→1280 no h-scroll (1.4.10); targets ≥24px (primary 44–48px); rendered-verified
G5
A6
Conversion, UX & content
one primary CTA/screen; ≥5-state inventory per interactive screen; trust signals at the decision point; FK grade ≤ ceiling
G4
A7
Visual & brand system integrity
primitive→semantic tokens, no hardcoded brand values; light+dark, no residue
G3
A8
Modern-stack engineering hygiene
semantic HTML, modern images/fonts, CI wired (LHCI + axe/pa11y), no secrets in bundle
G5
A dimension with any core-path failure caps at ≤ 50 (see each gate); a silent omission of an archetype-irrelevant sub-check scores as a miss, not an N/A.
Family B — this-pipeline yardstick (self-check). A pipeline that grades others should be gradeable itself. Self-scored against Family B — flagged as a self-grade, not an independent audit, with the honest residuals called out below the table:
Dimension
This pipeline
B1 Gate structure & terminal verdict
9 named fail-closed gates, one terminal Go/Conditional/No-go at G9.
B2 Falsifiable criteria
Every acceptance line in §4 anchors to a number or a named artifact (contrast ratios, px targets, KB budgets, CWV thresholds) — no bare adjectives.
B3 Correct agent dispatch
Every gate names one primary + any explicit co-owner, cross-checked against the real web-design roster (7 agents / 12 execution skills + this pipeline); SEO's technical/content split is named, not implicit; G5's static-stack build now dispatches the dedicated static-site-implementation skill (added this PR), not "the agent authors it directly."
B4 Dependency DAG & parallelism
§3's graph names 2 real parallel branches + the G8 content/technical split-edge, not a flat list.
B5 Escape hatches & seams
§6 names every seam + the one mandatory zero-exception rule + defer-not-duplicate; surfaces the unclosedecommerce-dtc seam instead of hiding it.
B6 Evidence discipline
Every numeric bar traces to a plugin-internal knowledge file / skill / template (durable on install — no gate depends on a dated research-scratch link); the llms.txt caveat is stated; the headline CWV gate is split so no pre-launch checkbox claims field data that cannot yet exist (G7a lab+budget+RUM is the mandatory bar, G7b field-data is a post-launch G9 Condition); volatile facts carry Last verified: 2026-07-06.
B7 Severity/owner/date
§5 defines one P0–P3 scale reused at every gate.
B8 Typed artifacts
§7 defines a per-gate typed-artifact ledger + federated sign-off + the plugin's Structured Output block.
Soft spots — three closed in this PR, one inherent (a self-grade, not an all-green feature list; the closes are surfaced, not hidden).
B5 — the ecommerce-dtc seam — CLOSED (web-design side). The platform-vs-headless → unit-economics handoff is now a named cross-reference in web-design/CLAUDE.md's "Adjacent plugins" seam (added this PR); the pipeline treats it as an explicit G1 sub-step with a documented degraded stand-in. Residual: the reciprocal ecommerce-dtc-side cross-ref is a recommended follow-up in that plugin (a one-line edit outside this PR's scope), so the seam is named from here but not yet echoed back — B5 is substantially closed, one-directional.
B3 — the static-stack build-skill gap — CLOSED. The plugin now ships static-site-implementation (added this PR) — the generic semantic-HTML + token-driven-CSS build skill for the G1 static-first default (Astro/11ty/Hugo/plain). G5's non-React static path dispatches it directly, so the primary archetype's recommended stack now has a dedicated operationalizing skill, not "the agent authors it directly." This row is now clean.
B3/B6 — security-reviewer per-agent coverage — CLOSED.security-reviewer is now named in all 7 agents' ## Escalation routes (the 5 that lacked it each gained an archetype-appropriate row this PR), so the §1 seam-rule assertion is true per-agent as well as pipeline-side. This row is now clean (the G9 gate was always the enforcement regardless).
B6/B2 — several pre-launch bars degrade to Conditional when browser tooling is absent (INHERENT, correctly designed — not closeable by a doc edit). G5 Tier 3, G7a Tier 2, and G8's Lighthouse Tier 2 all close Conditional (not clean) with no reachable Chromium/headless browser. This is a property of fully-sandboxed execution, not a pipeline defect: the tiered fallbacks make it a closeable Conditional with a named-reviewer condition (CI re-runs the real audit), never a silent wedge. It stays surfaced as an honest ceiling on a tool-less agent's verdict.
Net self-grade: strong on B1 / B2 / B4 / B7 / B8; B3 and B5 now clean (B5 one-directional); B6 clean except the inherent tool-availability ceiling (#4) — the three previously-Conditional dimensions are closed in this PR, with the one remaining Conditional being a correctly-designed property of sandboxed execution, not a gap.
Final checklist (machine-scannable — the pipeline's own hygiene gate)
[ ] G1 Archetype declared (marketing | web-app | ecommerce); perf + a11y budgets numeric; stack w/ rationale + ≥2 alternatives; ≥1 success metric + measurement; escalation flags raised
[ ] G2 5 IA artifacts (sitemap, URL taxonomy, nav spec, content model, redirect plan); sitemap shape justified; URL→template map
[ ] G3 Primitive→semantic tokens (W3C JSON); light + dark, no residue; contrast pre-checked ≥4.5:1 / 3:1; design-system-audit clean
[ ] G4 Real copy (no lorem); microcopy done; every interactive screen has ≥5 designed states (empty/loading/error/partial/ideal); trust signals at decision point; voice per style-guide do/don't table + reading level ≤ target FK grade; ≤1 primary CTA/screen; forms minimal + justified; funnel instrumented; keyword briefs drafted
[ ] G5 Stack-conditional build skill (React/Fluent→fluent-react-implementation; static Astro/11ty/Hugo→static-site-implementation; React skill N/A on static); semantic HTML; tokens consumed not defined; reflow 320→1280; targets ≥24px (44–48 primary); rendered-verified (Tier1 MCP / Tier2 project headless-dep; else Tier3 Conditional); no secrets in bundle; HTTP/2-3 + modern images; CI wired (LHCI + axe/pa11y)
[ ] G6 WCAG 2.2 AA: Tier A (automated+keyboard+human SR) = clean pass, else Tier B (axe + Playwright a11y-tree/ARIA dump + programmatic focus-order sim) = Conditional + named-human-reviewer condition; 6 new-at-A/AA SC (2.4.11/2.5.7/2.5.8/3.3.8/3.2.6/3.3.7) + 3 carried (1.4.3/1.1.1/1.4.10) verified; P0/P1 remediated; findings = P0–P3 + owner + date; AAA stretch (1.4.6/2.4.13/2.5.5) on primary flows
[ ] G7 Pre-launch (G7a, mandatory): Lighthouse ≥90 (Tier1 live; Tier2 static-proxy→Conditional if no Chromium) + CI byte budget (G1 budget governs; web.dev default floor <170KB critical path / TTI<5s where G1 set none) + RUM wired. Post-launch (G7b): CWV Good @p75 both form factors (LCP≤2.5s, INP≤200ms, CLS≤0.1) from field data — a standing G9 Condition for greenfield, enforced pre-launch only where field data already exists
[ ] G8 Non-skipping headings; valid JSON-LD (visible content; pre-deploy = syntactic stdlib-`json` parse: valid JSON + `@type` + required props per shape map; full schema.org shape conformance via Rich Results Test + social-card validators post-deploy at G9); Lighthouse SEO ≥90 (same Tier1/Tier2 ladder as G7a); canonical + robots + sitemap; crawlable w/o JS; AI-crawler policy set; llms.txt NOT gated; G8a opened alongside G3/G4 (closes at G4)
[ ] G9 All specialist sections signed; security headers via curl -I against staging/preview pre-launch (no unsafe-* CSP), prod re-check + live Rich-Results/social-card confirmations recorded as standing Conditions; no pre-consent tracker fire; terminal Go/Conditional/No-go recorded
--- seams (verify fired where in scope) ---
[ ] Web-app app build seamed → frontend-engineering; brand/UX/tokens/WCAG audit retained here
[ ] Ecommerce growth/economics → ecommerce-dtc (unclosed cross-ref; platform-vs-headless is an explicit G1 sub-step); themed build retained here (or headless → frontend-engineering)
[ ] Auth / sessions / payments / PII / untrusted input → security-reviewer (MANDATORY, zero exceptions)
--- discipline (every gate) ---
[ ] Every archetype-irrelevant sub-check recorded as N/A — <reason>, never silently dropped
[ ] Every claim cited-or-flagged; volatile facts dated; arithmetic rules have a runnable checker
[ ] Every finding: P0–P3 + owner + target date (one shared vocabulary)
[ ] Every gate emits a typed Structured Output block; waivers recorded, never silent
[ ] Verdict rule: every applicable dimension ≥90 = gold-standard; any <70 = Conditional + launch-blocking condition
Family A/B yardstick: both enumerated inline in §9 — the consolidated Family-A A1–A8 dimension table (built site) + the Family-B self-check (this artifact); each Family A bar is also restated within its owning gate's acceptance criteria (§4)
Provenance (one-time, not a live dependency — see the "Derived from" note in the Standards basis): a session research pass over external standards (w3.org / web.dev / developer.chrome.com / OWASP) + a house dispatch-map review, scored against a gold-standard rubric; the durable in-plugin capture is knowledge/gold-standard-website-references-2026.md
Ce SKILL.md est tres volumineux, SkillsMP affiche donc ici seulement la premiere section.Voir sur GitHub
G7 Performance
Full weight; image/font budget dominates (asset-heavy, JS-light).
Full weight; JS/interactivity budget dominates — INP is the metric most at risk, not LCP.
Full weight; image budget (product photography) and third-party-script budget (payment SDKs, reviews, chat) both matter — third-party-script-hygiene is load-bearing.
G8 SEO/AEO
Highest weight — organic discovery is frequently the primary acquisition channel.
Often partially N/A — authenticated in-app views are not indexable and are not graded on SEO; state it explicitly, don't silently skip. Marketing/landing pages in front of the app get full weight.
High weight; product/category schema (Product, Offer, AggregateRating JSON-LD) is the acceptance-critical addition beyond the base checklist.
G9 Pre-launch
Full weight; Legal/compliance section may be thin (no payments/PII beyond a contact form).
Full weight; security-reviewer sign-off is mandatory whenever auth/sessions/user data exist.
Full weight; security-reviewer sign-off is mandatory for payments/checkout/PII, and privacy/consent mechanics (§4, G9) are acceptance-critical given PII at checkout.
G7a pre-launch: Lighthouse ≥90 + CI budget + RUM wired · G7b post-launch: CWV field verdict (standing G9 Condition for greenfield); the dual-analytics pair is ONE inventory line but MEASURED — fail if it adds ≥50 ms TBT or regresses LCP (third-party-script-hygiene §8, never a blanket pass)
Team Lead + all 7 (+ security-reviewer where in scope)
→ templates/launch-checklist.md
G3–G8 all closed/waived
— (convergence)
Federated sign-off + terminal Go / Conditional / No-go
A11y target declared
AA
≥1 success metric named (conversion rate, task-completion, AOV — archetype-dependent) with a stated measurement mechanism, so G4/G7 can wire it.
Escalation flags raised now, not discovered later: web-app with substantial custom interactivity → note the frontend-engineering handoff boundary before G5; any login / auth surface → flag the auth-identity seam now (OAuth/OIDC flow, session/token strategy, MFA — reviewed before G5 builds it, not discovered at G9; resolved operationally like the ecommerce seam — dispatch auth-identity if installed, else a documented non-specialist stand-in stamped re-run via auth-identity once installed, with the auth-architecture decision recorded as a G9 Condition); ecommerce → schedule the platform-vs-headless decision before G5. If ecommerce-dtc is installed in the consumer project — resolved operationally per §6 (attempt the dispatch; if its agent/skill can't be resolved, treat it as not-installed), not by an assumed oracle — that decision is its call (3-yr payback / assortment / LTV:CAC). If it is not installed (a plausible split — the two are separate marketplace plugins), take the documented degraded path: web-architect drafts a lightweight platform-vs-headless memo (themed-platform vs. headless, the obvious cost/skill trade-offs it can judge) explicitly stamped non-specialist stand-in — re-run via ecommerce-dtc once installed, and the unit-economics call (payback horizon, LTV:CAC) is recorded as an open Conditions: item at G9 rather than silently assumed. The sub-step is never left undefined because its escalation target is absent.
Objective bar. Budgets exist and are numeric (CWV + perf budgets per knowledge/web-platform-capabilities-2026.md); archetype is unambiguous. A brief with adjectives-not-numbers ("fast," "accessible") fails.
Fail-closed. No archetype, no numeric budget, no stack-with-alternatives, or no success metric → cannot proceed to G2. Loop back.
Artifact.discovery-brief.md (from the design-brief.md template).
Fail-closed. A "sitemap alone" is decoration — missing any of the 5 artifacts, or an unclosed content model / URL taxonomy (both load-bearing inputs to G3/G4), loops back.
Artifact.ia-plan.md (from the site-architecture.md template).
Fail-closed. Hardcoded hex outside token files, a missing/broken dark theme, or unaudited drift → loop back before G5 consumes it. A "we'll clean it up during build" waiver is not permitted here — rebuild cost is why this gate exists.
Artifact.design-tokens.json + design-system-audit-report.md (the design-system-audit skill's own inline output format — scoring table + top-5 fixes; templates/design-system-spec.md is the companion spec template for the system itself).
Tier 2
One primary CTA per screen (at most two); forms reduced to essential fields, each extra required field justified (each costs ~5–7% completion).
Content model from G2 populated; conversion funnel defined with an instrumentation/measurement plan.
Redirect plan (from G2) finalized with real 301 targets before any page is retired; keyword / search-intent brief drafted per template (feeds G8a).
(Ecommerce) product-page completeness + funnel wired so ecommerce-dtc can read AOV / LTV:CAC.
Objective bar. Yardstick A6 — single clear CTA, minimal justified forms, all copy final. Placeholder copy or competing CTAs cap the gate.
Fail-closed. Lorem in a "final" mock, an unjustified bloated form, or an interactive screen missing any of its ≥5 designed states (empty/loading/error/partial/ideal) → loop back. A partial waiver is permitted for non-critical page copy (e.g. a blog post queued post-launch) but never for primary-flow copy (CTAs, checkout, error states) or for a missing loading/error state on an interactive screen.
Artifact.content-ledger.md (KKCR matrix + finalized microcopy table; from content-style-guide.md).
ravenclaude-core/visual-feedback-loop
Web app:seam to frontend-engineering — hand off the design system + tokens + a11y-verified patterns; they own component architecture, rendering strategy (SSR/SSG/RSC/CSR), state/server-cache, in-code bundle/CWV, React component craft. Do not duplicate their build. G5→G1 loop-back (the one backward edge): if their rendering-strategy call invalidates the provisional hosting/CDN chosen at G1 (e.g. an SSR/RSC edge/Node runtime vs. the static-first CDN picked at G1), loop back to G1 to re-commit hosting under a recorded waiver (owner: web-architect + reason) — this is the only sanctioned backward edge, scoped strictly to the hosting-decision-invalidated-by-rendering-strategy case.
Headless ecommerce:seam to frontend-engineering for the app-grade storefront; web-design retains PDP/cart copy + conversion-design.
Any auth/payments/PII surface:security-reviewer (mandatory).
Acceptance criteria.
Semantic HTML first (ARIA only where semantics can't express it); exactly one <main>, labels associated to inputs.
Tokens consumed, not defined in components — zero hardcoded brand values.
Reflow 320 → ≥1280 CSS px with no horizontal scroll or loss of function (WCAG 1.4.10); touch targets ≥ 24×24 px everywhere, primary controls at 44–48 px (Apple HIG / Material).
Verified against the actual rendered layout across form factors — not from CSS inspection — via a tiered path so the gate never silently blocks on an optional tool, but also never claims a reader that does not exist:
Tier 1 — a browser-driving MCP is present (chrome-devtools-mcp, which this plugin documents as recommended-not-bundled and security-reviewer-gated — CLAUDE.md §13, so commonly absent by default): screenshot + list_console_messages + lighthouse_audit — the model sees the pixels. Clean close.
Tier 2 — a project-local headless browser is a declared dependency (Playwright / Puppeteer already in the build's devDependencies): drive it via Bash to dump the rendered DOM + accessibility tree + computed box geometry (getBoundingClientRect) at each breakpoint and assert no horizontal overflow at 320 px. Clean close. This path is not provided by ravenclaude-core/visual-feedback-loop — that skill's driver.py is a pass/fail referee over evidence the agent already captured, explicitly "NOT a browser driver … cannot navigate Chrome, take a screenshot, or run Lighthouse", and its only structural reader is the Power BI/Tableau PBIR-JSON linter, not a web-DOM reader. Tier 2 therefore requires a real headless-browser dependency in the project, not any bundled web reader.
Tier 3 — neither an MCP browser nor a project headless-browser dep is reachable (no rendering engine at all): the rendered-layout check cannot be run, so G5 closes at Conditional with a standing named-reviewer Condition (owner: frontend-implementer + target date, §5) to run Tier 1/2 once a rendering engine is reachable — the same first-class "tool unavailable → closeable Conditional, never a silent wedge" treatment G7a/G8 use. A static proxy runs in this tier only (grep the built CSS for fixed pixel widths > 320 px on non-overflow-x:auto containers) as a weak reflow sanity check that does not upgrade the tier past Conditional.
Tier 1 or Tier 2 is a valid clean close for G5; Tier 3 is a Conditional close; only an un-checked build (no tier run at all) loops back.
HTTPS + HTTP/2/3, modern image formats (AVIF/WebP + fallback) with srcset/sizes + below-fold loading="lazy", font-display: swap/optional with subsetted self-hosted or preconnected fonts, critical CSS inlined, no render-blocking third-party above the fold, build pipeline with tree-shaking/minification/Brotli, tested 404.
Production build clean: no console errors, lint/type-check/unit tests pass, no secrets in client bundles (grep the built output, not just the source tree).
security-reviewer has reviewed before G5 closes if the build touches auth, sessions, payments, PII, untrusted input, or file upload — the design/implementation, not just the terminal G9 sign-off. Auth/session mechanics are costly to rework late (the same "audited after the build hardens → 10× cost" trap §8 names for a11y); land the review while the surface is still soft. In scope but unreviewed → G5 cannot close. (G9 remains the mandatory final gate; this makes the first touchpoint pre-hardening, not terminal.)
CI wired: Lighthouse-CI budget assertions + axe-core/pa11y lint — the one line that turns every downstream number from a manual check into an enforced gate.
Fail-closed. Horizontal scroll at a common mobile width, hardcoded values, a secret in the bundle, a build with no rendered-layout check run at all (Tier 3's static proxy is the minimum — absence of an MCP tool alone is not a fail, it drops to a Conditional close via the ladder above), no CI wiring (auditing an unenforced build is theater — it passes today and silently regresses tomorrow), or an in-scope auth/payments/PII surface security-reviewer has not yet reviewed → loop back. For the seam variants, G5 passes only when the receiving plugin returns its build and the web-design a11y/perf bars (G6/G7) still gate it.
Artifact.build-report.md (CI run links, bundle-size table, secrets-scan result).
plus the 3 carried-forward SC most likely to regress in a rebuild
1.4.3
1.1.1
1.4.10
(The count is 9 verified SC — not "9 new in 2.2." The genuinely-new-in-2.2 set also includes three AAA criteria — 2.4.12, 2.4.13, 3.3.9 — which are not in this AA floor; of those only 2.4.13 is adopted as a stretch goal in the AAA line below, per the standards basis, which does not recommend site-wide AAA.)
Findings carry P0–P3 severity + named owner + target date with the SC citation.
P0 + P1 remediated before launch; skip link visible on focus, correct <html lang>, no color-only signifiers, prefers-reduced-motion honored.
AAA stretch adopted on primary flows only (not the whole site — W3C itself doesn't endorse site-wide AAA): 1.4.6 (7:1), 2.4.13 (focus appearance), 2.5.5 (44×44 px targets).
Fail-closed. Any open P0 (never waivable — includes CAPTCHA-only auth and contrast failure on body copy), or an open P1 with no recorded Conditions waiver → launch-blocking, loop back to frontend-implementer. An open P1 may instead take a Conditions waiver at G9 (owner + target date, §5) for a 🟡 Conditional launch; P2/P3 findings may take a waiver rather than blocking launch. (One waiver rule, identical to §5 and G9 — never a per-gate reinvention.)
RUM / Web-Vitals reporting wired now (at G7, pre-launch) so field confirmation starts the moment traffic arrives. (Ecommerce) funnel instrumentation live so unit economics are readable.
G7b — post-launch field-confirmation sub-gate (tracked as a standing G9 Condition, NEVER a pre-launch G7 blocker on a first launch):
CWV Good at p75 on both mobile and desktop from field data (CrUX/RUM) — not averaged across contexts: LCP ≤ 2.5 s · INP ≤ 200 ms · CLS ≤ 0.1. For a brand-new site this is recorded at G9 as a mandatory post-launch Condition (owner: performance-engineer + target date, via §5's P0–P3 scale), confirmed by the RUM wired in G7a. For a re-platform / live-site re-cert where field data already exists, G7b is checkable pre-launch and is enforced then.
Objective bar. Yardstick A2; §2 of the CWV knowledge basis. A green one-off lab run with no enforced budget and no RUM wired caps A2 at ~75; any CWV in "Poor" (once field data exists) caps ≤ 50.
Fail-closed. Pre-launch: a Tier-1 Lighthouse lab < 90, no CI-enforced budget, RUM not wired, or "we check perf manually" → loop back (all measurable before launch). A Chromium-less agent does not loop back on the unrunnable Lighthouse bar — it closes Conditional via Tier 2 (static proxy + named-reviewer Condition), the same no-silent-wedge rule the rest of the pipeline uses. Post-launch: any CWV Poor at p75 → the standing G9 Condition trips, remediate. A first launch is never blocked on field data that cannot yet exist, nor permanently stalled on a Lighthouse run it cannot execute; it is blocked on the budget+RUM bar it can meet today.
Artifact.perf-audit-report.md (Lighthouse scores, budget-assertion CI link, post-launch field-monitoring plan; from performance-budget.md).
OG/Twitter card tags present + well-formed in the emitted <head>
Lighthouse SEO ≥ 90
same Tier-1/Tier-2 ladder as G7a
(Ecommerce:
Product
Offer
AggregateRating
)
AEO proxies (checkable, weighted lower): content server-rendered/crawlable without JS execution; a deliberate allow/deny of AI-crawler user-agents (GPTBot / ClaudeBot / Google-Extended / PerplexityBot) in robots.txt; question-shaped headings with the answer in the first 1–2 sentences where content is genuinely Q&A-shaped.
llms.txt is low-cost insurance, NOT a gate — no major AI vendor has committed to reading it as of 2026. Do not penalize absence, do not over-credit presence.
Fail-closed. Invalid structured data (fails the local syntactic JSON-LD parse pre-deploy — malformed JSON, missing @type, or a required property absent per the shape map; the live Rich Results Test provides full schema.org shape conformance post-deploy at G9), accidental production noindex, or heading-rank skips → loop back to the owning half. Web-app: authenticated in-app views are marked N/A — behind login, not indexable by design, never silently omitted; marketing pages fronting the app still gate at full weight.
Security headers present + non-default (verify with curl -I): HSTS (2-yr, includeSubDomains, preload); a real CSP with no unsafe-inline/unsafe-eval in script-src + a frame-ancestors directive; X-Content-Type-Options: nosniff; Referrer-Policy: strict-origin-when-cross-origin; explicit Permissions-Policy; the COOP/CORP/COEP trio. Pre-launch this runs against a staging/preview deploy — a production URL does not resolve before go-live, so curl -I against the not-yet-deployed domain is a DNS/connection failure, not a header failure; a post-launch re-check against the real production response is recorded as a standing Conditions: item (owner + target date, §5), mirroring the G7b field-data pattern. If no staging/preview URL exists either, verify the header set against the deploy/CDN configuration that will emit it and make the live-URL check the post-launch Condition.
Live third-party post-deploy confirmations deferred from G8 — the Rich Results Test, Twitter Card Validator, and Facebook Sharing Debugger (all JS-driven UIs needing a public URL) — are recorded as standing post-launch Conditions: items (owner: web-architect + target date) run against the staging/preview or production URL. The Rich Results Test here supplies the full schema.org shape conformance the pre-deploy syntactic parse (G8) deliberately could not — the offline parse only proved valid JSON + required-property presence per the shape map; the live tool confirms the vocabulary is used correctly and the rich-result is actually eligible. They do not block Go (the offline parse already gated G8 pre-launch on syntax/presence); they close the loop on the live shape check + OG-card rendering once a public URL exists.
Consent mechanics verified (if trackers/cookies present): non-essential trackers do not fire before consent (first-load network trace, no interaction); granular per-purpose toggles; "Reject all" at equal prominence to "Accept all" in the same layer; no pre-ticked boxes; symmetric (same-click-count) withdrawal.
Every checklist item is independently falsifiable — a number or a pointable artifact, never "looks good."
For a greenfield first launch: the CWV field-data confirmation (G7b) is recorded as a standing post-launch Conditions: item (owner performance-engineer + target date), with RUM already wired (G7a) to close it — it does not block Go, because field data cannot exist pre-launch. (Skip only if field data already exists — a re-platform/re-cert — in which case G7b was enforced pre-launch.)
Terminal verdict recorded: 🟢 Go | 🟡 Conditional | 🔴 No-go, with a Conditions: field populated when Conditional.
Objective bar. Security headers (the HSTS/CSP/COOP-CORP-COEP set, encoded above and in templates/launch-checklist.md) + consent mechanics — the consent-UX dark-pattern checks (equal-prominence reject, no pre-ticked boxes, symmetric withdrawal) are operationalized directly as the falsifiable sub-checks in G9 + templates/launch-checklist.md (GDPR / EDPB dark-pattern guidance), while the consent-gating / script-loading side (consent-mode v2, default-denied loading, don't-fire-before-consent) is backed by third-party-script-hygiene; yardstick A4. Any pre-consent tracker fire, dark-pattern banner, missing CSP, or secret in bundle caps A4 ≤ 50 and is a No-go/Conditional.
Fail-closed — exactly one of three states, never a silent trail-off.
🟢 Go — every section signed, no open P0/P1, security-reviewer signed where applicable.
🟡 Conditional — a Conditions: field is required, listing each open item with owner + target date (§5). A single sub-70 rubric dimension caps the launch here. Launch proceeds; conditions tracked to closure.
🔴 No-go — an open P0 (never waivable), an open P1 with no recorded Conditions waiver, or a missing mandatory security-reviewer sign-off, blocks launch outright. Loop back to the failing gate; this pipeline does not permit launching around it. (A P1 with a recorded Conditions waiver is a 🟡 Conditional launch, not a No-go — that is the one waiver path, and it never applies to a P0.)
Artifact. The completed launch-checklist.md, ending in ## Final sign-off.
