Shelter
The promise: governed agents that feel protected.
The ARCT field guide · three steps
This is the practical version of the process behind ARCT: decode a reference, create a sequence of controlled video moments, and turn them into one smooth, reversible website.
save this page — the prompts are ready to paste ↓
Do not begin by copying the page section by section. First identify its interaction grammar—what is pinned, what scrubs, what changes, and why—then map that grammar onto a real landing-page narrative.
The promise: governed agents that feel protected.
The product’s core capability, isolated and inspected.
A real dashboard showing performance improve.
Every decision becomes visible and explainable.
The assistant diagnoses a failed agent run.
Create a goal for this project: Build a complete cinematic, scroll-scrubbed SaaS website from the supplied reference recording and generated media. Act as the project manager. Break the work into research, story, asset planning, implementation, performance and QA. Keep a checklist and maintain a handover file containing every decision, media path, completed step and unfinished task. Continue until the site has been tested in the browser on desktop and mobile. Ask only when approval is genuinely required, and do not mark the goal complete while required work remains.
Build an original cinematic launch website for a fictional SaaS product called ARCT, short for Agent Runtime Control + Telemetry. ARCT is a control plane for production AI agents.
Use the attached website and complete screen recording only as interaction references—not as a design to copy. First audit the full forward and reverse experience. Identify the pinned chapters, camera or object movement, scroll timing, media transitions, text choreography, cursor behaviours and UI overlays. Then translate that interaction grammar into five landing-page chapters: a governed shelter hero, a floating runtime-artefact feature scene, a Control Plane product dashboard, a trace tunnel proof section and a NIVA incident-investigator problem section, followed by a return to the opening world.
Keep the visual media full bleed. Layer real semantic HTML product interfaces over it. Use restrained polar art direction, reversible scroll motion, fog transitions and cursor-reactive details. Keep accessibility, reduced-motion behaviour, mobile composition and smooth reverse scrolling in scope. Before coding, return a chapter map with the purpose, visual beat, copy beat, scroll range and transition for every scene.
A reliable cinematic page is not one impossible AI video. It is four or five short, controlled clips with designed overlaps. Generate matching endpoints first, then ask Flow to make only the movement between them.
Before writing code, divide this experience into five short, reversible video chapters. For every chapter, give me: (1) an exact first-frame image prompt, (2) an exact last-frame image prompt, (3) one Google Flow first-and-last-frame motion prompt, (4) a negative continuity lock, (5) an approved filename, (6) the recommended duration and active motion window, and (7) the transition into the next scene.
Keep the palette, camera, materials, lighting and object identity consistent across every chapter. Use one continuous shot with linear motion and no cuts. Do not ask Flow to morph unrelated objects or scenes. Hide scene changes with designed fog, blur and bloom transitions in the browser.
Use the attached master image as the exact world, material, lighting, lens and colour reference. Original polar-tech art direction, not another brand and not a copy of its model. Cinematic 16:9 desktop composition. Nearly monochrome blue-grey palette, soft overcast light, frosted mineral surfaces, restrained cold-white internal glow, fine airborne snow, subtle volumetric haze, distant low-contrast mountain forms. Camera and subject geometry must remain physically coherent. Leave clean negative space in the corners for HTML typography. No words, letters, numbers, logos, interface, watermark or border inside the image.
One continuous shot. Linear motion at constant speed. No cut, dissolve, time jump, speed ramp, camera shake, handheld movement, lens change, object replacement, new object, melting, liquid morph, duplicate geometry, extra panels, text, logo or watermark. Preserve the exact material, proportions, landscape and exposure from the supplied frames.
Orbit slightly and separate only the existing upper panels.
The camera performs a slow, mechanically smooth 18-degree orbit to the right with a very small dolly out. At the same time, only the existing upper shelter panels separate radially by a few centimetres, revealing a cold-white inner light. The foreground ridge produces subtle parallax and fine snow drifts laterally. All movement begins immediately and remains linear until the last frame. Add the negative continuity lock.
Rotate one rigid object while the camera stays locked.
The exact mineral monolith rotates clockwise at constant speed while drifting upward in a straight line. The camera remains locked. A very faint scan light passes once across the surface from left to right; fog moves slowly behind it. The object must remain rigid and retain every edge and cavity. Add the negative continuity lock.
Make the system visibly organise itself as the user scrolls.
Exact top-down locked camera. The six existing arc segments rotate and slide along the circular track at constant speed until they form one complete ring. The centre particles contract into a compact core while the outer rim gradually illuminates clockwise. Rigid mechanical motion, no bouncing and no overshoot. Add the negative continuity lock.
Reveal the assistant through particles, then keep the final pose stable.
Locked camera. The existing cold-white particles gather inward at constant speed and resolve into the exact arctic assistant shown in the final frame. Preserve the platform, lighting, silhouette and facial proportions. End with a calm, fully formed pose and no residual camera motion. No walking, waving, facial change or body transformation after formation. Add the negative continuity lock.
The page uses ordinary HTML, CSS and JavaScript. GSAP ScrollTrigger supplies the master scroll timeline; the videos are prepared for accurate seeking; product copy and UI remain crisp HTML and SVG above the footage.
Find the approved 1080p video clips and copy them into this project's media folder. Preserve the original source files. Inspect their duration, resolution, frame rate and keyframe spacing. Create separate web-scrub MP4 derivatives with a keyframe on every frame using this command, replacing the input and output names as needed:
ffmpeg -i INPUT.mp4 -c:v libx264 -preset medium -crf 18 -g 1 -keyint_min 1 -sc_threshold 0 -pix_fmt yuv420p -movflags +faststart -an OUTPUT-web-scrub.mp4
Build the complete page in plain HTML, CSS and JavaScript using GSAP and the ScrollTrigger plug-in. Create one pinned 100dvh master timeline. Map scroll progress to each video's currentTime; do not play the videos normally. Use Blob URLs when the host does not expose usable byte-range seeking. Give every chapter a deliberate active-motion window followed by enough hold time for the copy and UI to be read.
Crossfade chapters through fog, blur and bloom instead of hard cuts. Keep the logo, copy, dashboard, trace cards and controls as semantic HTML or inline SVG above the videos. Every animation must reverse correctly when scrolling upward. Add loading, mobile and prefers-reduced-motion fallbacks. Test multiple forward and reverse scroll checkpoints in a real browser and fix console errors, missing media, jumping playheads and slow frames before calling the build complete.
Polish the final interaction layer. Add a cursor-local video reveal using a second copy of the active video under a feathered radial CSS mask. Let the secondary video advance independently with GSAP while the pointer is present, but never let pointer movement change the master scroll time. Add a visible cursor cue and a coarse-pointer fallback.
Use the 21st.dev Stacking Cards component as a visual and mechanical reference, then adapt the behaviour to the existing GSAP master timeline without adding React or Motion. Before the final assistant explanation, introduce two larger diagnostic cards: a healthy-run comparison and a policy-threshold panel. Animate their inline SVG paths with stroke-dashoffset, opacity and transforms. Let each card remain readable, then scale and offset it into a three-card stack. Make the entire sequence reversible and validate it on desktop and mobile.
Use this 16:9 version in your video or save it as the quick visual recap inside your own project brief.
The prompts above contain the public-facing brief, production constraints and QA standard. The private skill stays private while it improves across more builds.