Pangram verdict · v3.3
We believe this text is mainly human-written, with some AI content.
AI likelihood · overall
HumanArticle text · 1,332 words · 4 segments analyzed
One notable difference between working on Claude and the other productivity tools I've touched is that feature requests are surprisingly not the #1 discussion topic. Instead, people want to know how I'm using Claude, how they should be using Claude, and whether there are any cool magic spells I can share.In this post, I'd like to walk you through the process of redesigning my homepage without touching or running code or even using traditional Claude Code. No PRs were made, nor did I open GitHub. If you haven't seen the full end result, maybe check it out first. 0:00 /1:31 The redesign in 90 seconds Roughly 60 threads worked, often in parallel, on my increasingly outrageous requests, which included creating a CGI parody of a Werner Herzog documentary in Blender and synthesizing soundtracks from scratch. Despite the complexity, a considerable amount of that orchestration happened on my phone rather than my computer - in fact, I didn't have to run any code locally. Claude worked entirely in the cloud.I'm writing this in late September 2026, and I suspect my workflow will look very different in six months. Even compared with the beginning of the year, I'm noticing a stark difference - my work is now a lot more creative, less bounded by the technologies I've mastered, and so much more visual and emotional. I've used Opus 5.5 for all the work.I started by creating a new Claude Project. Projects are becoming a layer that spans Chat, Cowork, and Code; you can think of them as sitting above domain-specific concepts. You talk to one Claude with the ability to spin off additional sessions, share files and memory between them, and supervise the entire fleet of Claudes to achieve goals.Earlier this year, I would have joked that I get a lot done by telling a bunch of Claudes to do work. Now, I've put one Claude in charge of each of my projects, and in turn, that Claude knows how to manage tasks, investigations, and threads. We're talking more about goals and less about how to achieve them.My little Project setup.Every thread gets its own computer in the cloud, where it installs whatever it needs (Blender, FFmpeg, Playwright), checks its work with screenshots from a headless browser, and publishes a private preview page for me to click through. All threads share one folder of files and one memory. Claude is smart enough to write down decisions I made in memory and makes sure newly started threads follow them without me having to repeat myself.Brainstorming a new directionI started the way I'd start with a design team: by dumping a bunch of inspiration on the table. Those included 90s rave posters, movies I liked, ads that spoke to me, photos I keep coming back to, even just concepts I like.Ten threads explored those directions in parallel with very little guidance. One came up with portfolio directions for somebody who really likes camping in national parks; another imagined what an 80s Ghost in the Shell website would look like. Both, it turned out, felt a little overdone to me. I wanted to make something new, not just clone something already out there.Whenever I liked a direction, I asked Claude to double down, get creative, and give me five more variations of the same direction. Twenty of the 25 brainstorm sites. Top left: the VHS direction that became the room.While most directions used HTML, Three.js, and other libraries to achieve some of my hopes and dreams, I pretty quickly found myself pushing Claude to use Blender. Claude doesn't click around Blender's interface, instead writing code and using Blender as a Python module. When I briefly considered turning my website into a cozy campsite, Claude promptly modeled one. Some of Claude's first Blender models for an abandoned idea (a cozy camp fire)Given the direction of vintage technology I liked (Game Boys, record players, Walkmans, etc) and the task of coming up with a creative portfolio gallery, Claude eventually suggested VHS tapes.Claude's prototype for a portfolio where every gallery item is a VHS tapeTogether, we combined Claude's idea (VHS tapes) with my own vision for the website, which included an interactive version of the kind of room a 90s German movie would give a journalist.The VHS direction's second versionClaude's second draft, including my directions about the room setup and encouragement to use Blender creatively, was shockingly good. The version I shipped took roughly 60 iterations, but the second draft, which arrived 39 minutes after my first brief, is already recognizably the room you see today.Early on, Claude decided that the code should be modular, allowing threads to work on elements of the room in parallel. As an example, each tape is its own file that paints onto a 2D canvas. export default { id: "notion", title: "Notion", years: "2023–2025", mount({ canvas, ctx, width, height, audio }) { return { frame(t, dt) { /* draw one frame */ }, input(event) { /* play, stop, next, prev, click */ }, stop() { /* clean up */ }, }; }, };After every iteration step, Claude published a new version of an artifact, a private web page where I could see the current version, both on my computer and my phone. I constantly asked for a few options, clicked through them, sent back some critique, and went on with my day. Doing that from my phone felt particularly powerful. I would be on my bike, have some ideas while riding, quickly fire off some ideas to Claude, and check out the results on the next coffee stop.Building the roomLet's talk about the room, which is probably where I pushed Claude to be most ambitious. One of the most beautiful parts of LLM tools is that they allow me to do things I wouldn't have been able to do before - I have no idea how to use Blender. I describe the things that I want, Claude writes the geometry as code, renders a still, and I give feedback.The room is now ~3,000 lines of Python code, controlling Blender as a library.
Some of those scripts are:tape_model.py (332 lines): one VHS cassette, 187 × 103 × 25 mm, as a hollow shell with reels, tape ribbon, guides, and a hinged dust door.gear90s.py (659 lines): a 90s journalist's desk. A phone with a microcassette answering machine standing on a phone book, a Rolodex, floppies, a pager, a solar calculator, a Discman, and a Walkman, with cords running behind the desk to a wall socket.retro_pc.py (431 lines): a late-80s beige PC clone from a made-up brand, with two 5.25" drives, a key lock, a red paddle switch, a 14" green CRT on a tilt-and-swivel foot, and an XT-style keyboard.room_back.py (290 lines): the half of the room behind the camera, which you only see in reflections and in free-look: a four-panel door with a key in the lock and light under it, a Bakelite light switch, a stucco cornice, and a tiled stove.Claude decided on its own when to scour the Internet for freely licensed models and when to model props from scratch.
Some furniture came from Poly Haven (CC0-licensed), but most items are "made by Claude" - like the desk, the TV & VCR, Discman, Walkman, the entire old PC, or the entire room geometry. The surface textures are code, too: Claude's wood, plaster, plastic, and metal are procedural noise, not photographs.The home view in Blender.I kept pushing for a photorealistic look, which soon had us talking about light maps.
Real-time lighting in a browser can't do the soft bounce light I wanted, so Claude decided to use Cycles, Blender's path tracer, and bake the light into textures. It's a surprisingly involved process: Every static surface gets a second UV set packed into atlases, and each atlas gets up to four maps: its color without light, the light from the room's own lamps (desk lamp, city glow, moon), the light from the TV alone, glowing plain white, and the room with the desk lamp switched off.