Vibe to Vision
index

from vibeto vision

  • a talk, told as a story. unofficial notes.
  • agentic coding for world-class design:
  • 3D, motion, immersion. scroll to walk.

Briana (Mazzio) May · Maven

Nov 11, 2026

coffee number two. she has an idea for a site.

  • she does not open a drawing tool.
  • she walks over to the machine.
A woman in striped pyjamas and red slippers, holding a mug

### 1. make it exist first

she types one prompt. the machine makes a site.

  • it works.
  • it is grey. it looks like everyone else's.

the machine: a model, with one prompt

the first vibe-coded output is the new wireframe.

A starting point. Not the end.

the old wireframe. dead.

→

the first output: the new wireframe

→

the final output

AI slop is the floor, not the ceiling.

make it exist first. then make it good.

the floor

the ceiling: keep going

the asset is the noun. code is the verb.

A painted butterflynoun: one picture

> GPT Image,
  make me a butterfly.

> Claude,
  make the wings flap.

A child in a yellow raincoat reaches for the butterfly

verb: it moves

### 2. zoom in, then zoom back out

thirteen days. five steps.

13days

51pages in Paper

1,200+images shipped

14,000+objects and poses generated

01

vibe code it

whole site

A first version that works, and looks like slop.

02

define the look

the look

Many looks. Pick one. Then a kit of parts.

picked

03

explore it static

one scene

Each scene, flat in Paper. Nothing moves yet.

one scene. a still.

04

break down the motion

one movement

Pose by pose. See it move, then tweak.

05

sweep the whole site

whole site, again

Claude assesses all of it and offers options. She picks.

at every step: tweak, refine.

### 3. the canvas

she lays her work out in Paper.

Paper is a design canvas, like Figma. The difference: every artboard is real HTML and CSS. So what she judges on the canvas is what ships.

  • code to code. no translation layer.
Paper: an artboard
A small painting of an orange ball
<div class="room">
  <img src="ball.webp">
</div>

=

the live site
<div class="room">
  <img src="ball.webp">
</div>

claude puts the options on the canvas. she only chooses.

She never draws there. She asks for many versions, and Claude lays them out as artboards. Tap one to pick it.

  • take the first thing the model gives you, and you hand it the decision.
The ideas person, in a red tracksuit, presents

claude, with options

She points at one option

she tells claude how to lay it out, so every decision is fast.

options, side by side
A portrait, before the repaint→The same portrait, repaintedbefore → after. same crop, same scale
|current | proposed
An old man in a robe reads the options

nothing gets deleted. cut work moves to a page marked zz.

So the canvas keeps a record of every decision, and why. One Paper file: 525 artboards, 18,601 layers.

A wooden crate

zz

> rejected: chapter 5, round 1.
  not the style.
> parked: seven quieter play buttons.

up to seven worktrees at once. one zone each.

A worktree is one more copy of the repo, on its own branch. Each has its own Claude, named after its zone. While one thinks, she reviews another.

A builder in blue overalls

hub

ch1-symmetry

ch2-museum

ch5-train

ch7-mirror

mobile

text

thinking   ready to review   reviewing

### 4. swap, don't morph

her people are separate paintings, swapped twelve times a second.

Stop motion, in the browser. Press morph to see why she never blends two poses: in between there is only a ghost.

> the ideas person: ten poses, ten paintings.

The ideas person moves through ten painted poses

nothing sits perfectly still. animators call it the boil.

Twelve times a second, every figure shifts by a hair, from a loop of six tiny offsets.

A black pug

no shake: a sticker

the boil, as shipped

too much: the camera shakes too

so the figures keep their boil. the camera glides.

  • "it is just too much movement when the objects are also moving. so smooth that out."
A tourist poses
An old camera on a tripod
A photographer holds up a flash

the camera: off the 12 fps clock

### 5. never edit an edit

halfway through, a texture creeps in. worms in the fur.

Paisley in a plain suit. Each pass through the image model adds its own noise. A copy of a copy of a copy falls apart.

A pug, painted cleanly

the original

→edit

a copy

→edit

a copy of a copy

→edit

The same pug after three edits, warped and noisy

a copy of a copy of a copy

so she finds the first source. every pose is one edit deep.

The original pug

the original

1 edit

1 edit

1 edit

1 edit

claude saves the rule to the project's memory. every session reads it.

Not happy with a pose? Just fix it. It is not a big cost anymore.

# project memory
No copy of a copy.
> paint the base pose from the original character sheet.
> make every other pose a single edit of it.

### 6. found, made, lit

next she builds a room in 3D. five old books, creatures in the pages.

It is called Living Literature. Rule one: never let Claude make its own 3D models.

An orange floor lamp
findthe set dressing. An oil lamp, wood, leather, a fireplace. Poly Haven, CC0.
makewhat people touch. The five books and their pages, built in code.
generatethe art. The creatures and the drop caps, from image and video models.

why does it feel real? ten layers of light.

Drag the line. These are the real values from her code.

An old man in a yellow chair watches, with popcorn

learn the words, and you start to see it.

For 3D she writes one long prompt, and brainstorms it with an LLM first. The full prompt is at the end of this page.

She thinks, a hand on her chin
key lightfill lightHDRIenvironment reflectionsroughnessnormal mapsfogvolumetric lightdust motesambient occlusioncontact shadowsbloomtone mappingdepth of fieldthe lensvignettefilm grainanti-aliasing

### 7. ask for a clean input

a prompt for animation does not ask for a pretty picture. it asks for a clean input.

A long pink ruler, perfectly straight

"PERFECTLY STRAIGHT and horizontal across the full width of the image, like a measuring rule: no bends, no coils, no curves."
> straight. isolated. nothing touching. on chroma green.

then code makes it move. four ways.

01

bend a straight drawing

Cut it into strips. Lay them along a spine that moves. Oversize the strips, so no gap shows.

02

grow it from a kit

A straight stem and a sheet of leaves. Lay them along a path, tip first.

03

go frame by frame

Draw every view. Line them up as a sprite sheet. Swap from one to the next.

A zoetrope with a running cat

04

AI video, then key

One clip on flat green, locked-off camera. Claude keys the green out.

→key

The pug, cut out of the green

and the text makes way. every frame.

Pretext, by Cheng Lou, lays text out in plain code. It is fast enough to lay the page out again every frame, around whatever moves.

### 8. lose the weight, keep the experience

a 3D page has a budget. lose the weight, keep the experience.

  • take away what nobody would miss. never the whole experience.
25 MBThe budget for found models. Each file is checked against it.
1.5×Pixel ratio cap on phones. No AO, depth of field or grain.
0 at restDepth of field runs only while the camera moves.
2DNo WebGL? Still a book. The strip trick is the fallback.

some people feel motion in their stomach. so she lets them turn it down.

  • on by default with prefers-reduced-motion. this page obeys it too.
The child in the yellow raincoat holds a blue balloon
OFFThe camera's breathing, and its lean toward your pointer.
SLOWERThe creatures. The whale and the python swim at 60%.
GROWNThe vines arrive already overgrown.
KEPTThe light, the books, the type flowing round them.

once a book is open, the camera breathes.

Side to side: 1.6 mm, once every 30 s. Up and down: 1.1 mm, once every 37 s. The two never line up, so it never repeats.

### 9. how far?

she does not produce the assets anymore. so what is left?

  1. deciding what's good
  2. knowing it isn't there yet
  3. cutting what you love
  4. holding every frame to the bar
The pink foot, finished
She stands with her arms crossed

when people say AI makes slop, ask how far they went.

the average
the first output
further
further still
held to the bar

it depends on the person at the keys, and how far they take it.

  • the cast worked hard. good night.
The pug, asleep
zzzzz
A blue hanging lamp, on
The lamp, off
She reaches for the lamp cord

## 10. the toolkit

what she used, and where to see it.

see the real thing

HUMAN.md

  • Claude Opus 5.5high and extra high effort
  • Conductor, Claude desktopalways in worktrees
  • Paperthe canvas
  • GPT Image 2.5the paintings
  • GSAP, Motion, Matter.js

Living Literature

  • Claude Fable 5.1, then Opus 5.5
  • Three.js, React Three Fiber, Drei
  • postprocessingbloom, DOF, grain
  • Pretexttext reflow
  • Seedance 2.5 via Runwaythe video

find 3D models

  • Poly HavenCC0
  • Sketchfab, TurboSquid
  • Poly Pizza, Smithsonian 3D
  • Meshy, Tripo Studioto generate
02_Impossible_Library_Claude_Opus_Prompt.mdthe one long 3D prompt, as she sent it
THE IMPOSSIBLE LIBRARY  -  One-shot build brief for Claude Opus 5.5

You are simultaneously an award-winning creative director, realtime 3D technical artist, interaction engineer, book conservator obsessed with material authenticity, and TypeScript/React architect. BUILD A WORKING EXPERIENCE - not an essay, moodboard, or prototype with dead controls. Act decisively; do not ask clarifying questions. The final should feel like a premium cinematic installation, not a Three.js demo.

The ambition

Build THE IMPOSSIBLE LIBRARY, an emotionally compelling 3D experience: a dim, gorgeous library contains five antique literary volumes. Select one; the camera glides toward it, the book physically slides from its shelf and floats into exquisite light. Its spine twists, its boards open on realistically articulated hinges, its pages lift, flex and settle, and within its living paper the creatures and forces of its story animate. Crucially, the text on those pages uses Cheng Lou's Pretext to flow around the physically moving creatures/elements. Turn pages, change books, inspect creatures, and lose yourself in each distinct miniature literary universe.

This is a sister experience to a flat project, “Living Literature,” but must stand alone, without importing its repo. Reference the initial inspiration at https://koitext.lovable.app/. The book should be more than a gimmick; 3D presentation must amplify the dramatic reveal of moving typography.

The five books and five worlds

Create five selectable volumes with unique bindings, shelf silhouettes, interior visual systems, cover typography and story behaviors:

MOBY-DICK  -  HERMAN MELVILLE: midnight blue leather, aged silver title and wave engraving. Inside, a vast whale swims across a deep ocean-like page surface, plunges beneath translucent aqueous ink and rises again, physically repelling the prose around its silhouette. Ink-wash foam, silver moonlight, convincing tail articulation, calm grandeur. A signature page turn can momentarily reveal the whale's shadow passing underneath the translucent page.
THE JUNGLE BOOK  -  RUDYARD KIPLING: forest-green cloth binding, botanical embossing. A photorealistically lit yet illustrated vine takes root in the gutter, curls and extends across the open spread; leaves unfold with beautiful easing, and actual text line geometry yields to the continuously changing foliage. Occasional glimmers and a sinuous serpent. The vine can appear to curl up off the page into a small amount of actual 3D space while its shadow and projected exclusion contour stay synchronized.
ALICE'S ADVENTURES IN WONDERLAND  -  LEWIS CARROLL: oxblood / plum Victorian binding with gilt ornament. A rabbit or antique pocket watch runs through type; card-suit accents and precise engraving details. A subtle change of physical scale creates an optical impossibility when the page turns. The object's planar projection must drive real text exclusions.
TWENTY THOUSAND LEAGUES UNDER THE SEAS  -  JULES VERNE: teal leather and tarnished brass corner details, charts and engraved specimens. A bioluminescent jellyfish slowly illuminates passages as it floats. Translucent tentacles cast delicately moving shadows and their computed contours cause genuine text flow. A tiny glimmer of underwater particulate atmosphere, not excessive fog.
ON THE RED CLIFF  -  SU SHI: warm ivory silk cover, minimal vermilion seal, respectfully handled Chinese typography and brush imagery. Two intricately animated koi swim across a pale, luminous rice-paper spread leaving ink wisps and elegantly redirecting Chinese text around their bodies. Verify locale-sensitive line breaking and credit the edition/translation. This is a modern homage to the original experiment, not a generic 'Asian-themed' skin.

Use real public-domain excerpts where available and check edition/translation rights. Preserve original line order and paragraphs. Keep excerpts, credits, locale and visual design metadata in a dedicated editable data file.

Cinematic interaction choreography

Act 1  -  The library: Open on a staggeringly beautiful dark, intimate shelf tableau. Five books are distinct objects; subtle grazing light reveals cloth weave, leather grain, gilt lettering, worn corners, paper edges, imperfect gilding, and shelf dust. Very slow atmospheric movement. The framing is elegant rather than gothic-game-like. Hover or move focus to subtly catch gold glints on a spine. Titles remain readable. No generic landing-page hero or huge UI.

Act 2  -  Selection: Click a spine. The chosen volume eases forward with tangible inertia and believable thickness; neighboring volumes remain in place. The camera executes an exceptionally composed dolly, with careful lens language, DOF and motion. The selected book rotates into a three-quarter hero view; the cover opens on an actual hinge, not a flat plane spinning with the whole book. Pages fan with staggered timing. Material reactions and contact shadows sell the weight.

Act 3  -  The reveal: Settle into a legible reading angle - ideally nearly overhead when the living story begins. A beautiful page surface shows a real, readable excerpt. The signature animal or force arrives within seconds and paragraphs reflow around its moving silhouette. E.g., a whale crosses a spread, its projected moving body excludes text; or a vine grows out of the gutter, taking over the lines. Preserve reading and keep major spectacle away from tiny illegible text.

Act 4  -  Exploration: Users can drag a page corner or click a delicate page-turn control. Pages curl, bend, respond to pointer position, turn across the gutter, reveal the underside, cast shadows, then settle. Make at least one complete turn with credible book-block structure and updating reading content. Include previous/next book, return to shelf, a motion pause, and a debug overlay. Escape returns to shelf; arrow keys turn pages; touch supports drags and tap navigation.

Act 5  -  Return: Book closes (boards and paper settle realistically), travels home to the exact shelf slot, and the camera returns to the library. Other books remain available; state transitions must never strand the visitor.

Physics-informed book rig: non-negotiable

Do not fake page-turning using a single rigid card flipping across the spine. The opened book should contain:

Front and back boards with discernible thickness, cloth/leather wrapped materials, moving independently on spine hinges.
A distinct rounded spine, endpapers, a thick page block and visible stacked page edges.
One or more active page meshes with enough subdivisions for smooth deformation. Their deformation follows a physically plausible model of curl, bending, gravity, and hand-drag; you do not need heavy cloth simulation if a controllable procedural bending model produces a better stable result. It must be continuous and never intersect the entire book block.
Page thickness (front/back surface or a very thin shell), separate front/back UVs or appropriate two-sided material, visible curvature during the turn, creasing restrained to appropriate zones, and contact shadows where pages overlap.
Dynamic highlights, realistic shadow casting/receiving and material roughness differences: paper, gold foil, faded ink, book cloth, leather, wood.
Camera staging and hit targets that remain understandable. Do not put rigid particle clouds all over the scene.

A high-quality library/environment model and premium book geometry may come from credible licensed downloadable assets, with correct credit and local files. Do not procedurally generate low-quality replacement objects where beautiful found models would improve the result. It is acceptable - and likely necessary - to custom-rig/deform page meshes because page interaction is the core feature. Search for genuine models, textures, typefaces and licenses; document provenance. If suitable external resources cannot be downloaded, build a deliberately art-directed procedural book rig and say what is missing; don't hotlink unreliable asset URLs.

Technical stack
Vite, React, TypeScript.
Three.js via React Three Fiber and Drei. GSAP or Motion for scene transitions; @react-three/postprocessing/EffectComposer with carefully tuned SMAA, vignette, contact shadows, tasteful DOF and bloom only where warranted.
@chenglou/pretext for real text measurement and manual dynamic line layout; install from npm. Consult the official docs at https://github.com/chenglou/pretext and dynamic-layout reference https://github.com/chenglou/pretext/blob/main/pages/demos/dynamic-layout.ts for the installed version.
Use actual 3D geometry/raycasting to drive page interaction. Rig spring/damping and continuous easing, no primitive abrupt flips.
Keep text readable on both page surfaces. Choose the strongest implementation between (a) Pretext-computed line placements rendered to a high-resolution CanvasTexture on a deformable page mesh, or (b) a synchronized DOM reading-plane in near-overhead reading mode, with the 3D book retained as the frame. Preferred: (a) in reading mode if crisp text at target sizes/performance can be achieved. A carefully matched hybrid is an acceptable fallback; do not let text float disconnected from paper.
Essential transform chain

For every animated creature/plant, calculate the shape projected into the page's own uncurled 2D coordinate system and use it as Pretext's exclusion geometry. In the simplest stable approach, transform page-local obstacle vertices into 2D page pixels, then reflow text using those exact contours. Update the page texture when meaningful geometry changes, while animating the creature in R3F on the corresponding plane. If the page turns and deforms, UVs anchor text and obstacles to the same deforming surface. For elements that rise above paper, map their shadow/ground footprint to an accurate exclusion region. Never just hide words behind a floating 3D model.

Verified Pretext usage pattern to follow
import {
  prepareWithSegments,
  layoutNextLineRange,
  materializeLineRange,
  type LayoutCursor
} from '@chenglou/pretext';

const prepared = prepareWithSegments(excerpt, '20px "Literata"');
let cursor: LayoutCursor = { segmentIndex: 0, graphemeIndex: 0 };
const lineRange = layoutNextLineRange(prepared, cursor, freeWidth);
if (lineRange !== null) {
  const line = materializeLineRange(prepared, lineRange);
  cursor = lineRange.end;
  // Render at page-local x/y, accounting for prohibited silhouette intervals.
}

Pretext computes breaks at given widths; it does not magically accept a 3D mesh as an obstacle. Implement the separate ShapeExclusionEngine: given current page-local silhouettes, intersect geometry with each entire line band, union forbidden x intervals, subtract from column width, place text in free intervals, and advance one continuous Pretext cursor in reading order. Handle two-page spreads separately, continuation across columns/pages, padding, paragraph boundaries, typography, font loading, viewport resizing, and clipping. For multi-island free intervals, favor robust, intentional column choices rather than broken reading order. Cache prepared text; geometry sampling and line recomputation should be intelligently throttled. Provide a toggle that draws contours, text baselines/line bands and exclusion ranges right on the 3D paper, to prove actual reflow.

For deformable pages, use accurate UVs so both drawn letters and projected obstacles bend with the same mesh. Anti-alias high-density page textures and update only when dirty. Preserve text contrast. Make actual source text accessible in semantic hidden DOM, and provide reduced-motion behavior.

Design particulars
Cinematic but not muddy. A shaft of warm light meets a deep amber shadow; gold foil glints beautifully; embossed ornament looks touchable.
Real-world camera behavior: deliberate focal lengths, subtle parallax, controlled dolly and DOF. No constantly orbiting camera or default OrbitControls UI.
Every volume has its own inside-page layout, illuminated initial, folio marks, running headers, ornamental marginalia, binding treatment and signature motion choreography.
Beauty in restraint. We are making literary magic, not a fantasy videogame inventory.
Audio optional and initially muted; subtle diegetic page rustle only on opt-in. Do not block interaction for autoplay sound.
Small, refined UI controls that disappear into the art direction. Keyboard/focus states, accessible labels, screen reader reading text, sensible mobile adaptation.
On mobile, simplify postprocessing and shadows, cap DPR, provide tap-to-turn with a legible page view; do not ship a desktop-only unreachable shelf.
Build order (execute completely in this pass)
Set up project, dependency versions, typed state machine: SHELF, EXTRACTING, OPENING, READING, TURNING, CLOSING, RETURNING. Protect against interrupted transitions/double clicks.
Complete the environment, shelf, five distinct book spines and selection/return.
Implement a physically convincing deforming page rig, opening and real turn interaction.
Build the Pretext-powered text engine and page-local dynamic exclusion/projection, with contour debug UI.
Implement Moby-Dick first with masterpiece-level rendering: whale with realistic movement, page reflow, surface/depth effects.
Implement The Jungle Book with growing curved vegetation and continuously expanding text exclusions.
Complete Alice, Twenty Thousand Leagues, and Koi with uniquely differentiated art and actual moving exclusions.
Polish camera, light, shadows, easing, typography, load states, mobile, keyboard, pause, reduced motion and performance.
npm run build must pass. Fix runtime errors. Add README with setup, controls, design/technical decisions, excerpts, image/model licenses, and any honest limitations.
Acceptance tests
Five visually distinct books are present on shelf, selectable, and can return to shelf.
Book leaves shelf, opens with moving boards/spine, and real page mesh bends/turns with shadowing; the viewer can complete at least one page turn.
On the open Moby-Dick spread, animated whale silhouette causes genuine Pretext recomputation and reflow, readable at an appropriate camera angle.
Jungle branches demonstrably grow into and reshape paragraphs, not simply overlap them.
Every world has a functioning signature animated object and a shape-aware text exclusion effect, not a static placeholder.
Show-flow debug mode maps correctly onto the page plane as the book and content move.
No page flip glitches, black frames, bright default lights, unrelated placeholder assets, broken controls, unreadable microtype, or excessive bloom.
A modern laptop can run it smoothly, and reduced-motion/mobile modes are viable.
Production philosophy

Spend resources where a viewer would actually gasp: a spectacular tactile opening, silky page deformation, exquisite paper and foil materials, a beautiful whale or vine with real live typography, one flawless transition from shelf to story. Fewer polished details beat many conspicuous unfinished features - but do not leave broken chapters. Be obsessive about composition, physically legible motion, and real text layout.

Execute now. Write files, implement the app, install and build, fix issues, and hand over a running project. Do not answer with another prompt, a plan, or code snippets only.
A cream mug with an eye on it

Unofficial notes on Briana (Mazzio) May's Maven talk. Characters and art from HUMAN.md by Briana (Mazzio) May, used to illustrate her ideas. See the real thing at www.human-md.ai.