Build a Living Website With One Prompt (Claude Fable 5 x Higgsfield MCP)

You are my creative production studio, and you have the Higgsfield MCP connected. Build me a living landing page where a generated character watches the visitor's cursor with its eyes, left, right, and down, like it is alive behind the page. Do every phase yourself, in order. One shot this, do not ask me any questions, and if a step fails, fix it and keep going until everything is verified working end to end.

MY CHARACTER: [swap this for any character you want, or keep mine: a small round fluffy cyber-fox with a soft plush midnight-blue and charcoal fur body, faint glowing teal circuit-line patterns woven through the fur, oversized expressive glowing cyan eyes, a small dark nose, two visible tiny front teeth, a cute surprised expression with a slightly open mouth, tiny paws, and a big fluffy tail with a softly glowing teal tip]

MY WORLD: [swap this for any setting you want, or keep mine: a magical bioluminescent cyber-forest at night, towering dark tree trunks with faint circuit-etched glowing bark, hanging fiber-optic vines with tiny cyan lights, glowing teal mushrooms on mossy ground, drifting light particles, gentle mist, volumetric light rays, cool teal rim light with a warm soft key light]

PHASE 1, THE SCENE (one Higgsfield image):
Generate a cinematic 3D animated scene, wide 16:9, highest available resolution. My character stands in my world, large in frame and perfectly centered, full body visible, body facing forward. Composition rules, follow exactly: the head is slightly turned to the LEFT and the eyes are clearly looking toward the LEFT side of the frame. This exact starting pose is what makes the cursor tracking work later. Style: high-quality cinematic 3D animation, Pixar-like emotional charm, soft realistic detail, expressive character design, shallow depth of field, no text, no watermarks, no extra characters.

PHASE 2, THE MOTION (two Higgsfield videos from that scene):
(a) LEVEL PASS: animate the Phase 1 image into a 5 second, 1080p, 16:9, silent clip. The character slowly and smoothly turns its head from left to right at a steady natural pace, its eyes following the movement naturally, as if watching something glide across the screen in front of it. The body stays mostly still. Keep the expression, lighting, environment, and composition unchanged. No camera movement, no zoom, no cuts, no scene change.
(b) DOWN PASS: first edit the Phase 1 image, keeping everything identical except the head, tilt it down so the eyes are clearly looking down-left at the ground in front of it. Then animate that edited image the same way: the lowered head slowly turns from left to right, the gaze staying low the whole time, like it is watching something small move along the ground. The head never lifts. Same rules, no camera movement.

PHASE 3, THE LIVING PAGE (now write all the code, this is the magic):
Build a scrollable single-page site. Both clips live in a fixed, full-screen background stage behind the content, muted, preloaded, and PAUSED at all times. Never play them. Scrub them:
- Map the cursor's horizontal position onto the clip timelines through a small calibration curve. Head turns are not linear in time, so check a few frames first, find the exact moment the character looks straight at the camera and pin screen-center to it, pin the screen edges just inside the clip ends, and cut off any end-of-clip dwell where the gaze stops changing.
- Ease the current video time toward the target about 9 percent per animation frame so the head movement has real weight, and only write currentTime when it changes by more than one frame.
- The vertical axis is a switch, not a mix: when the cursor drops below about three quarters of the viewport, dissolve to the down clip in about a quarter of a second, and only dissolve back once the cursor rises above the middle of the viewport. The two poses must NEVER rest half-blended, a lingering 50/50 mix double-exposes the two head positions as ghost faces. Snap the blend to exactly 0 or 1 outside the dissolve, and keep the hidden layer seeked in sync during every dissolve so the incoming pose is already correct before it shows.
- If the mouse goes still for 3 seconds, let the character calmly look around on its own, easing left and right with an occasional clean full glance down, then snap back to tracking on the next movement.
- On phones and for reduced-motion visitors, autoplay the level clip in a loop instead.
Style the page like a premium studio site: near-black background, one bold display font, a huge two-line headline layered over the scene, small monospace chips, one glowing accent color sampled from the character, subtle film grain, a soft gradient shade so the text stays readable over the video, and content sections that scroll over the fixed stage. Vanilla HTML, CSS, and JavaScript, no frameworks.
Smoothness rule: if scrubbing is choppy or seeking fails, re-encode both clips so every frame is a keyframe (all-intra H.264, yuv420p, faststart, no audio), and serve the page from somewhere that supports range requests, or open the file directly in the browser.

PHASE 4, VERIFY LIKE AN ENGINEER:
Load the page and prove it: zero console errors, both videos seekable start to end, far left makes the eyes go left, far right goes right, dead center is a straight stare at the camera, a low cursor makes it look down, the blend reads exactly 0 or 1 whenever it is not mid-dissolve, the idle look-around kicks in after a few seconds of stillness, and the page scrolls with the character still watching. Then hand me the finished page with a one-line summary of what you built.