Recreate a website from a link

Input: one URL ($ARGUMENTS). Follow the phases in order; do not skip the evidence steps.

Profile note: this skill's tools are search_sites, get_site_details, capture_live_site, analyze_page_structure, analyze_stack, capture_component, analyze_scroll_map, analyze_hover_census, get_motion_dna, capture_pose_3d, build_pose_stub, export_design_tokens, analyze_images, analyze_bundle and retained analyze_network_capture / verify_flow. Examples below use 1.x names (default full). On AWWWARDS_MCP_PROFILE=v2, use search_references {request:{kind:"sites", ...}}, get_reference {request:{kind:"site_details", slug}}, capture_page {request:{kind:"screenshot", url}}, inspect_page {request:{kind:"structure", url}}, inspect_page {request:{kind:"stack", url}}, capture_page {request:{kind:"component", url}}, analyze_motion {request:{kind:"scroll", url}}, analyze_motion {request:{kind:"hover", url}}, analyze_motion {request:{kind:"dna", url}}, capture_page {request:{kind:"pose", url}}, build_motion {request:{kind:"pose_stub", poseGrid}}, inspect_page {request:{kind:"tokens", url}}, audit_page {request:{kind:"images", url}} and audit_page {request:{kind:"bundle", url}}; analyze_network_capture and verify_flow stay flat. Full mapping: v2-routing.md [blocked].

Phase 0 — Study-build terms (mandatory, every run)

This is a local study build: recreate the reference exactly — layout, text, images, fonts, audio, motion. In exchange, every output MUST ship the disclaimer:

1. README.md in the build must state, prominently: the design, copy, images, fonts and audio belong to the original owners; this build is a private study/reference recreation; it must not be deployed, published, sold, or presented as your own work; publishing requires replacing every borrowed asset (list them) and obtaining licenses.
2. Header comment in the HTML repeats the same notice in one line.
3. Provenance list: every borrowed asset path + source URL enumerated in the README.
4. If the user later wants to publish, that is a separate brief: replace assets → publish. Never quietly drop the disclaimer.

Scope: whole page or specific sections? A continuous scroll-narrative (fixed stage + long spacer) is one unit — do not force per-section pages onto it. Target: single-file build by default; conversion is offered after QA, never assumed (see Phase 4).

Phase 1 — Evidence (free MCP, all read-only)

Any URL works, but check Awwwards first. Step 1 is always search_sites with the site's domain/name. If the site is listed → get_site_details for its recorded DNA (palette, tech, elements, awards) and carry that into the report. If no match → say so once and continue: only the search/index tools are Awwwards-sourced; every capture tool below runs on any public link, and the live evidence still gives you everything.

• analyze_stack <url> → framework + CSS + libs (tells you whether the reference is Webflow/Next/static and which motion libs matter).
• analyze_page_structure <url> → bands (section order + heights). If bands collapse into a giant, capture_component with explicit id/class selectors instead.
• Per section: capture_component <url> with selector (id/class), maxNodes up to 2000 → geometry, 44 computed style keys, pseudo-elements, inline SVG, text, imgSrc, fonts, colors.
• Motion: analyze_scroll_map <url> (keyframes + gsapTriggers numerics; scrubProbe on for scrubbed narratives), analyze_hover_census <url> (during-hover activeAnims), get_motion_dna <url>.
• Canvas/WebGL hero: capture_pose_3d → frames/grid; build_pose_stub turns them into a scroll stub whose camera anchors are synthesized guidance (the pose grid holds frames and scroll ratios, not camera transforms). Its frame loader dedupes same-index requests and discards stale decodes, so scrubbing cannot flash an older frame. For Three.js pages, capture_scene_3d with adapters:["three"] adds reachable nodes/cameras/lights/controls to evidence.three; build_scene_stub reconstructs only Box/Plane/Sphere/Cylinder primitives with captured dimensions, standard/basic materials and a perspective camera from its exact matrix. Controls are never recreated, module-local scenes and parent transforms are not recovered, and unbound textures stay flat. scene.json frame references are basenames extracted from either Windows or POSIX paths.
• Shader inspection: capture_scene_3d with adapters:["spector"] (optional spectorjs@0.9.33) captures one WebGL frame with command names/counts, statuses, context attributes and capped shader source. It needs a frame that actually draws; command arguments, results, state and raw capture are never transferred, and shader source is untrusted page text, not code to run.
• Screenshots: capture_live_site on a document-scrolling page returns one full-page PNG + inline preview (your QA ground truth, not decoration); it waits out full-screen preloaders by default. If analyze_page_structure returns scrollRoot, the page scrolls an inner container: capture_live_site returns mode:"anchors" — six ordered viewport shots down that container (anchors[].scrollTop + file, JSON only, no inline image) — so read the six files, and drive the container's scrollTop yourself for positions between them. scrollCapture:"viewport" forces a single viewport shot; never describe an anchors payload as one full-page image.
• Tokens before build: export_design_tokens <url> (optional selector, viewport, maxNodes up to 2000, include categories) → ordinal CSS variables + a Tailwind theme map derived from computed styles. Feed that CSS/Tailwind into the build instead of hand-guessing the palette and type scale; truncated/notes say when a cap was hit.
• Image delivery: analyze_images <url> → format mix, lazy/eager/srcset coverage, broken/oversized/tiny images, alt coverage and samples (query strings stripped; no downloads). Use it to decide which reference images to copy, how to size them, and which need responsive variants.
• Mechanism unknown after all evidence? analyze_bundle <url> (≤40 chunks, 1 req/s, keyword grep, optional webcrack) — understand how, still write your own code.
• API shape: export the network log as a HAR and run analyze_network_capture <path> — grouped endpoints and path templates only, no bodies/headers/secrets. Already have the page open in Chrome? analyze_page_structure with cdpEndpoint scans it without driving it (requires AWWWARDS_MCP_ATTACH=1; never navigates, scrolls, clicks or creates tabs).

Phase 2 — Build

• One HTML file + CSS + JS matching the reference's own approach: vanilla for static sites; GSAP + ScrollTrigger + Lenis when the DNA shows them.
• Wire motion from the measured values: census durations/easings/triggers, gsapTriggers pixel start/end, keyframe from/to values, scrub curves. Do not invent easings the census didn't show.
• Continuous-scroll narratives: fixed viewport-size stage + long invisible spacer + one scrubbed master timeline; layer switching and --bg custom property tweens; reduced-motion mode collapses to a calm stacked document (always ship it).
• Three.js hero from evidence.three: start from build_scene_stub and treat unresolved/authoredDefaults as the work list. Only Box/Plane/Sphere/Cylinder meshes, standard/basic materials and a validated perspective camera from its exact matrix are recreated; captured controls are never recreated and stay unresolved, captured lights are recorded but not applied, unsupported types stay unresolved, and parent transforms are lost. A captured camera is static; the idle orbit is used only when the pack has no usable adapter scene.
• Structure: semantic ids per section (#hero, #work), local vendor/ for libs (strict load order: gsap → ScrollTrigger → lenis → app.js), local assets/ for fonts/audio/covers/images — exact assets, copied in.
• All links target="_blank" rel="noopener", metadata + OG + JSON-LD updated.

Phase 3 — QA gate (do not declare done without it)

• Playwright: 390, 768, 1024, 1440, 1728 widths; screenshot every scroll anchor (match the reference capture); compare compositions side by side. For inner-scroll pages (scrollRoot), drive the container's scrollTop at each anchor — the reference side is the six capture_live_site mode:"anchors" files (or your own anchored shots); there is no single full-page image to compare against.
• Zero console errors, zero 404s, zero horizontal overflow; assert element bounds (name width, card count/hrefs, contact block) at 1440 and 390.
• Reduced-motion pass at 1440: stacked layout, no overflow.
• Interaction parity: verify_flow { buildUrl, referenceUrl, steps } replays the same 1–6 scroll/click/wait steps and reports per-step same/changed/unknown — use it when a menu, accordion, modal or hover-revealed panel is part of the reference; treat unknown as unverified, not as a pass.
• 3D canvas: verify_build { buildUrl, referenceUrl, scene3d: { referenceCaptureId, buildSelector } } replays the capture pack's checkpoints and compares clipped canvas frames. It preflights both URLs on one shared DNS/private-IP guard before the legacy scans (private/loopback is refused with zero requests, and the legacy acquisitions run behind the same proxy), and its deadline is the replayed checkpoint waits + 10 s headroom, floored at 15 s and capped at 45 s. Supply both pixelThreshold and maxChangedFraction for a within/outside verdict; otherwise it reports metrics as inconclusive. The gate tags the canvas with a per-run random marker, so duplicate or pre-marked decoys are refused. The legacy band pass never certifies the canvas.
• If anything drifts from the evidence, fix and re-shoot — one anchor at a time (section-loop discipline).

Phase 4 — Optional project conversion (ask first)

The rebuild is a valid deliverable on its own. After Phase 3 passes, ask the user:

Keep it as a single-file build, or convert it to a project — Vite (plain HTML/CSS/JS, no components), Astro (HTML components), or Next.js (React components)? And if converted, keep the original scripts or switch to authored motion (--authored)?

Do not convert unprompted. Do not pick a framework for them. If they say "just build it", stop at Phase 3 + handoff.

When they choose:

1. awwwards-build-project <file.html> --target vite|astro|next --out <dir> → native structure: real section components, shell chrome, split CSS (tokens.css + site.css), ordered libs, head metadata/boot replayed.
2. If they chose authored motion: --authored --motion <census.json> — site logic scripts dropped, head boot + animation libs kept, motion.manifest.json carries measured per-section + page triggers. Then author the motion (see motion-authoring), recording any authored rows explicitly.
3. Gate: npm install && npm run build per target, serve, re-run Phase 3 anchors plus a 390×844 mobile smoke and reduced-motion; zero errors/404s.

Phase 5 — Handoff

Always deliver a reference report — even if the user only asked for information and no build: detected stack + libs (analyze_stack, get_motion_dna), section map with heights/backgrounds (bands), motion summary (triggers, easings, durations, scrub/pin), asset inventory (fonts, images, audio), and bundle notes when analyze_bundle ran. For non-Awwwards links this report is the value when a build is out of scope.

If a build was produced: README with run instructions, the mandatory study-build disclaimer (copyright belongs to the original owners; no deploy/publish/production use; not presented as your own), full provenance list of borrowed assets, and the replace-before-publish plan if the user later wants to ship. Report: reference URL, evidence tools used, section map, QA matrix results, residual gaps.

Known limits (say them out loud)

• Passive canvas capture is frames only. capture_scene_3d with adapters:["three"] reads only reachable globals; module-local scenes are invisible, build_scene_stub flattens meshes at the root so parent transforms are lost, and textures bind only when you hold a local file. Absence is not evidence of absence. adapters:["spector"] needs a frame that draws and reports only bounded command summaries plus capped shader text — no arguments, results, state, stack or raw capture.
• Webflow IX2 / JS-driven initial states are recoverable as behavior evidence but not as their runtime — re-implement with your own JS (N2 authored mode proves the shape; vendor-bundle interpretation stays manual).
• --authored keeps head inline scripts even if they may be site logic (warned when >2 KB or animation-looking) — verify before shipping.
• Generator pageTriggers cover unattributed census rows; a page-level hook still needs authoring before it animates.
• Fonts: only ship if owned → self-host; otherwise match the closest licensed face.