AI COWORK ® STUDIO
————————————————
** AI PRODUCTS **
————————————————

DIRECTING TO AI THROUGH DESIGN

An AI-assisted workflow for building interactive design systems

SCROLL TO DISCOVER

00 — Overview

AI Workflow

AI Implementation Web Design Motion Frontend
Type
Self-initiated
Timeline
Jan 2026 – Ongoing
Role
Frontend Designer
Tools
Claude, Codex
Three.js, GSAP

01 — Problem

Define Problem

I built this because I kept running into the same problem. The tools exist, but using them well is harder than it looks. Most people interact with AI through a chat box. I wanted to find out what happens when you push it further: use it as a production partner, not a search engine.

But the real question I was trying to answer was: how do you give an AI taste? How do you build a shared aesthetic, label it, describe it precisely enough, so that what comes out actually matches what you had in mind? This site is the experiment. Every piece of it was built without writing a line of code by hand. What I was figuring out the whole time was the language. The exact way to describe a visual feeling so the AI can replicate it perfectly. That is what this project documents.

02 — My Workflow

How I Work

AI Workflow Process
01
02
03
04
01
Ideation
REFERENCES Studied motion-focused and 3D web experiences before writing any code. Identified specific techniques and visual languages worth adapting for this project.

DIRECTION Assembled two moodboards to lock in the visual direction early. One focused on material and form, the other on typography and graphic system. Both fed into the final visual language.
02
Experiment
STRUCTURE Started with a rough page structure to test the scroll flow. Sections were blocked in quickly to find the right rhythm before any visual polish was applied.

INTERACTION Multiple approaches to the 3D scene and card interactions were tested in parallel. Speed mattered more than quality at this stage.
03
Test
REBUILT The workflow layout was rebuilt after the first version felt wrong on screen. The pinning behavior broke completely and required a full rewrite. The hero scroll animation produced a visual glitch and the entire approach was replaced.

REVERTED Two versions of the camera movement were tested. Both felt wrong once seen in the browser. The original position was locked. A fix for the hero headline created a worse problem than the original bug. Both fixes were reverted.
04
Refine
SPACING Typography scale and section spacing were adjusted repeatedly. Small changes in padding and proportion made large differences to the overall feel of the page.

MOTION Scroll speeds, easing curves, and transition timings were dialed in last. Each value was adjusted until the movement felt intentional at full scroll speed.

03 — AI Limitations

When It Breaks

AI
Cannot See
No Aesthetic Sense
Memory Resets
Does Not Self-Audit
HUMAN
Needs To See It First
Describes With Feeling
Changes Direction Mid-Build
Does Not Always State The Boundary

04 — What I Found

What
I Found
1
Method 01: Create Reference
In Practice
I described the hero background as "warm, soft, like paper." Claude returned a generic cream — nothing close to the Figma file. Replacing the description with the exact hex #f4f2ef produced a match on the first attempt.
#01
2
Method 02: Set The Boundary First
In Practice
I asked Claude to adjust the card layout. It also shifted camera.position from (0, 5.0, 8.66) without being asked — the cinematic angle was lost. Adding "do not change camera.position" to every subsequent prompt stopped the drift.
#02
3
Method 03: Commands Not Explanations
In Practice
"Animate the 3D objects rotating slowly" produced chaotic spin on each axis. Rewriting as "Three knots orbit at 120° apart, one full revolution every 8 seconds" generated the hero animation directly on the first attempt.
#03

06 — Design

Design

At first, I considered exploring the project through industrial design and physical product development. As the idea evolved, however, I realized that I was more interested in how a concept could be communicated through 3D, motion, and interactive web experiences.

I therefore developed the project as an experimental, AI-assisted website. This direction gave me an opportunity to explore how 3D motion, scroll interaction, and cinematic language could be translated into a digital experience.

Before building the website, I studied a range of 3D and motion-focused websites and created several moodboards to analyze their camera movement, materials, typography, composition, and pacing. This research helped me establish the project's visual direction and explore how AI could support the process from concept development and prototyping to implementation.

VISUAL DIRECTION — MATERIAL & FORM
Moodboard: chrome materials and 3D forms
VISUAL DIRECTION — TYPOGRAPHY & GRAPHIC SYSTEM
Moodboard: industrial labels and typography
ORIGINAL AI DESIGN
Original AI card designs
I DESIGN AND REPLACED

07 — Tools

Stack
Claude
Codex
Three.js
GSAP
Figma

08 — Reflection

What I
Learned

Designing with AI changed what I thought the designer's job was.

What I originally thought

I assumed that if I wrote a detailed enough prompt, AI would be able to generate the website I had in mind.

What changed

During production,

Before the results became accurate, I had to define the design foundation myself:

  • typography weight, scale, and hierarchy
  • page spacing, card proportions, and information structure
  • 3D materials, lighting, and environment reflections
  • the camera movement from far to near and back into position
  • scroll, hover, flip, and loading interactions
  • which elements were allowed to change and which had to remain fixed
  • the references, screenshots, colors, and assets used as visual evidence

Only after these decisions were clear could Claude translate them into working HTML, CSS, JavaScript, Three.js, and GSAP.

What I learned

The main skill I developed was not simply prompting AI to write code. I had to specify:

  • where the camera should begin
  • how close it should move
  • how long the movement should take
  • how subtle the rotation should be
  • which easing curve to use
  • where the camera should finally settle

I also learned that generating more versions does not always improve the result. Knowing which parts are already working, what should remain untouched, and when to stop is also part of design judgment.

Going forward

For future AI-assisted projects, I would define the full experience before asking AI to build it:

  1. collect references and create a moodboard
  2. establish the typography, color, material, and layout system
  3. map the page structure and interaction flow
  4. prepare the 3D models, images, and other assets
View More Work