Made by mechanical means

One notable distinction between engaged on Claude and the opposite productiveness instruments I’ve touched is that function requests are surprisingly not the #1 dialogue matter. Instead, individuals wish to understand how I’m utilizing Claude, how they need to be utilizing Claude, and whether or not there are any cool magic spells I can share.

In this put up, I’d wish to stroll you thru the method of redesigning my homepage with out touching or operating code and even utilizing conventional Claude Code. No PRs had been made, nor did I open GitHub. If you have not seen the complete finish consequence, possibly check it out first.

Roughly 60 threads labored, typically in parallel, on my more and more outrageous requests, which included making a CGI parody of a Werner Herzog documentary in Blender and synthesizing soundtracks from scratch. Despite the complexity, a substantial quantity of that orchestration occurred on my telephone quite than my pc – in truth, I did not must run any code regionally. Claude labored totally within the cloud.

I’m penning this in late September 2026, and I think my workflow will look very totally different in six months. Even in contrast with the start of the 12 months, I’m noticing a stark distinction – my work is now much more artistic, much less bounded by the applied sciences I’ve mastered, and a lot extra visible and emotional. I’ve used Opus 5.5 for all of the work.

I began by creating a brand new Claude Project. Projects are becoming a layer that spans Chat, Cowork, and Code; you may consider them as sitting above domain-specific ideas. You speak to 1 Claude with the flexibility to spin off extra periods, share recordsdata and reminiscence between them, and supervise the whole fleet of Claudes to realize objectives.

Earlier this 12 months, I might have joked that I get loads accomplished by telling a bunch of Claudes to do work. Now, I’ve put one Claude in command of every of my tasks, and in flip, that Claude is aware of how one can handle duties, investigations, and threads. We’re speaking extra about objectives and fewer about how one can obtain them.

Every thread will get its personal pc within the cloud, the place it installs no matter it wants (Blender, FFmpeg, Playwright), checks its work with screenshots from a headless browser, and publishes a personal preview web page for me to click on by means of. All threads share one folder of recordsdata and one reminiscence. Claude is sensible sufficient to write down down choices I made in reminiscence and makes certain newly began threads observe them with out me having to repeat myself.

I began the best way I’d begin with a design group: by dumping a bunch of inspiration on the desk. Those included 90s rave posters, motion pictures I preferred, advertisements that spoke to me, pictures I hold coming again to, even simply ideas I like.

Ten threads explored these instructions in parallel with little or no steering. One got here up with portfolio instructions for any person who actually likes tenting in nationwide parks; one other imagined what an 80s Ghost within the Shell web site would seem like. Both, it turned out, felt a little bit overdone to me. I needed to make one thing new, not simply clone one thing already on the market.

Whenever I preferred a route, I requested Claude to double down, get artistic, and provides me 5 extra variations of the identical route.

While most instructions used HTML, Three.js, and different libraries to realize a few of my hopes and goals, I fairly rapidly discovered myself pushing Claude to make use of Blender. Claude does not click on round Blender’s interface, as a substitute writing code and utilizing Blender as a Python module. When I briefly thought-about turning my web site into a comfortable campsite, Claude promptly modeled one.

Given the route of classic expertise I preferred (Game Boys, document gamers, Walkmans, and many others) and the duty of developing with a artistic portfolio gallery, Claude finally steered VHS tapes.

Together, we mixed Claude’s thought (VHS tapes) with my very own imaginative and prescient for the web site, which included an interactive model of the form of room a 90s German film would give a journalist.

Claude’s second draft, together with my instructions in regards to the room setup and encouragement to make use of Blender creatively, was shockingly good. The model I shipped took roughly 60 iterations, however the second draft, which arrived 39 minutes after my first transient, is already recognizably the room you see right now.

Early on, Claude determined that the code ought to be modular, permitting threads to work on parts of the room in parallel. As an instance, every tape is its personal file that paints onto a 2D canvas.

export default {
  id: "notion",
  title: "Notion",
  years: "2023–2025",
  mount({ canvas, ctx, width, top, audio }) {
    return {
      body(t, dt) { /* draw one body */ },
      enter(occasion) { /* play, cease, subsequent, prev, click on */ },
      cease() { /* clear up */ },
    };
  },
};

After each iteration step, Claude revealed a brand new model of an artifact, a personal net web page the place I may see the present model, each on my pc and my telephone. I continuously requested for just a few choices, clicked by means of them, despatched again some critique, and went on with my day. Doing that from my telephone felt notably highly effective. I might be on my bike, have some concepts whereas using, rapidly hearth off some concepts to Claude, and take a look at the outcomes on the subsequent espresso cease.

Let’s speak in regards to the room, which might be the place I pushed Claude to be most formidable. One of probably the most stunning components of LLM instruments is that they permit me to do issues I would not have been capable of do earlier than – I don’t know how one can use Blender. I describe the issues that I would like, Claude writes the geometry as code, renders a nonetheless, and I give suggestions.

The room is now ~3,000 strains of Python code, controlling Blender as a library. Some of these scripts are:

• tape_model.py (332 strains): one VHS cassette, 187 × 103 × 25 mm, as a hole shell with reels, tape ribbon, guides, and a hinged mud door.
• gear90s.py (659 strains): a 90s journalist’s desk. A telephone with a microcassette answering machine standing on a telephone e book, a Rolodex, floppies, a pager, a photo voltaic calculator, a Discman, and a Walkman, with cords operating behind the desk to a wall socket.
• retro_pc.py (431 strains): a late-80s beige PC clone from a made-up model, with two 5.25″ drives, a key lock, a crimson paddle change, a 14″ inexperienced CRT on a tilt-and-swivel foot, and an XT-style keyboard.
• room_back.py (290 strains): the half of the room behind the digital camera, which you solely see in reflections and in free-look: a four-panel door with a key within the lock and light-weight below it, a Bakelite mild change, a stucco cornice, and a tiled range.

Claude determined by itself when to scour the Internet for freely licensed fashions and when to mannequin props from scratch. Some furnishings got here from Poly Haven (CC0-licensed), however most gadgets are “made by Claude” – just like the desk, the TV & VCR, Discman, Walkman, the whole previous PC, or the whole room geometry. The floor textures are code, too: Claude’s wooden, plaster, plastic, and steel are procedural noise, not images.

I saved pushing for a photorealistic look, which quickly had us speaking about mild maps. Real-time lighting in a browser cannot do the gentle bounce mild I needed, so Claude determined to make use of Cycles, Blender’s path tracer, and bake the sunshine into textures. It’s a surprisingly concerned course of: Every static floor will get a second UV set packed into atlases, and every atlas will get as much as 4 maps: its coloration with out mild, the sunshine from the room’s personal lamps (desk lamp, metropolis glow, moon), the sunshine from the TV alone, glowing plain white, and the room with the desk lamp switched off. The shader tints the TV’s mild with regardless of the tape reveals, so a blue tape turns the room blue:

vec3 lamps = pow(texture2D(irrA, uv).rgb, vec3(3.0)) * sA;  // desk lamp, metropolis, moon
vec3 television    = pow(texture2D(irrB, uv).rgb, vec3(3.0)) * sB;  // the TV alone, baked white
vec3 mild = lamps + television * tvColor;

Things that transfer (tapes, knobs, the VCR’s door, LEDs, the screens) aren’t baked. They’re lit reside.

Not to toot Claude’s horn right here an excessive amount of – however it’s insane to me that each one of this may merely occur in code.

For what it is value, whereas it is fairly cool to see Claude work within the cloud whereas I’m using my bike round, intense 3D work actually advantages from devoted GPUs. You can connect your individual pc in any thread by hitting the little “+” icon and deciding on Work Locally, which permits Claude to make use of your pc for speedier rendering.

The view exterior

I had very particular concepts about what the “exterior” ought to seem like: photorealistic, alive, with out consuming too many CPU/GPU cycles on guests’ machines. You’ll be unsurprised to listen to that when once more I did not actually must know what’d work greatest; we merely constructed each single possibility.

The first placeholder was a photograph (“Berlin Skyline” by Billie Grace Ward; CC BY 2.0). I assumed I’d be intelligent and generate a whole skyline in Blender. It seemed fairly good, simply not fairly pretty much as good because the photograph, so I deserted that route.

Claude had a greater thought: We used Depth Anything V2, an open depth-estimation mannequin, and turned the photograph right into a 50 KB depth map with two channels: a gentle one, so masts, indicators, and close to buildings slide in opposition to the skyline when the digital camera drifts, and a pointy one, so trains can disappear behind buildings.

That gave us the very best of each worlds – the seen photograph but in addition depth to allow animated lights and trains. They’re all drawn by the GPU on prime of the photograph, at virtually no CPU value:

Let’s discuss my favourite a part of the whole undertaking: making a parody of the “nihilistic penguin“, a scene from Werner Herzog’s Encounters on the End of the World: one penguin leaves the colony and walks inland, towards the mountains, alone.

Claude and I began easy, with a childlike animation drawn totally in SVGs. It allowed us to debate precisely what sort of parody I needed.

Once we had a fairly good thought of what I needed, I requested Claude to create a CGI model that will first mimic the unique carefully, in order that we may add my adjustments later. I’m advised that Claude did what a VFX artist would do:

• It broke the scene down into eight photographs, reduce on the unique’s shot boundaries.
• It tracked the digital camera body by body and tried to match each pan, zoom, and wobble of the hand held digital camera.
• It constructed (shock, shock) the global community utilizing code in Blender, as soon as once more utilizing the “Blender by way of Python” strategy. This contains every little thing you see – snow, mountains, gravel, penguins.

It rendered 2,787 particular person 640×480 frames at 23.976 fps, which resulted in a 116-second lengthy video.

Any good parody thrives on being fairly near the unique with just a few small particulars modified: a desk with a pc on the ice, a row of company flags throughout the lonely penguin’s approach (one for every place I’ve labored), or a foolish reference to the truth that I’ve a weblog that will get a brand new put up like yearly.

While Claude was total glorious at sound design, it initially tried some free open-source fashions to generate a voice over. It labored however was a little bit lackluster by way of high quality, so I signed up for an ElevenLabs Starter account and easily gave Claude the API key.

Oh, additionally, did I point out that Claude composed and synthesized three soundtracks on the web site from scratch? It used (once more, shock!) Python. It first wrote a rating, then a synthesizer engine. I’ve no thought how one can make music. I requested Claude about its strategy: “Deep Field is six four-bar chords in C# minor (C#m(add9), Amaj7(#11), F#m9, F#m(add9)/G#, C#m(add9), B) over a C# drone at 34.6 Hz. Late Shift is a 32-bar AABA tune type.” I’ve no thought what any of which means however I used to be capable of give Claude Rick Rubin-style suggestions within the type of “I like this” or “No, make it… much less excited”.

I feel loads of engineers are experiencing a disaster of religion: If I’ve received dozens of Claude threads buzzing away on the identical time, there isn’t any approach I’m studying all that code. Which is true: I’m not studying the code. I think that that is simpler to grasp for individuals who’ve labored each as an engineer and as an engineering supervisor – as a supervisor, it’s important to discover methods to determine high quality with out reviewing each single change your individuals make. Entire books have been written about how precisely you do this with out micromanagement, however I apply roughly the identical mechanisms. Two basic examples are demanding rigorous processes and measurements, or tasking particular person brokers with breaking issues (aka adversarial brokers who take into account their success to show that the code is dangerous).

Multiple QA threads drove the positioning spherical after spherical in a headless browser and wrote about 3,000 strains of findings to the shared folder. The coordinating Claude rapidly found out which different thread was at fault and handed the discovering on. The multi-Claude global community is, because it seems, not innocent. The adversarial brokers discovered cassettes with clean labels, leftover German, hyperlinks that stayed clickable behind playing cards, and a race if you change tapes rapidly. A single Claude thread does not write excellent code, a fleet of Claudes holding one another accountable meet my bar.

Hill climbing efficiency

Models are very good at hillclimbing. Tasking threads with particular targets, like body fee, draw calls, or shadow attracts, will get you very far.

You’d assume that fashions have a tough time telling whether or not two variations of a scene look roughly the identical. Claude got here up with a intelligent resolution – it took comparative screenshots and really calculated the pixel-by-pixel distinction. If you do this from a number of angles, you get a reasonably foolproof methodology of making certain that the efficiency positive aspects are actual positive aspects and never simply high quality degradations.

I wish to be delicate to those that could be studying this put up with an ache of their coronary heart, feeling prefer it’s outrageous that any person who by no means went by means of the ache of studying the speculation or apply of 3D modeling all of the sudden claims they’re constructing 3D issues. I’m certain you can discover single parts which are low high quality and name them slop, declaring variations between the factor I labored on for twenty-four hours and a room designed painstakingly by hand over weeks.

I’ve actually heard related emotions from engineers, each when Electron made desktop app improvement simpler and clearly through the rise of AI-assisted coding. I’m in all probability a reasonably median engineer, however I am knowledgeable one – I’ve labored on loads of issues that made it to many customers. In 2026, I would not have thought-about rebuilding my web site with out AI, however it’s humorous that I did not even attain for Claude Code as a devoted “for coding” instrument. I simply used Claude. Opus 5.5 and the generic harness did not have to be advised that we’re “coding” right here.

The San Francisco Museum of Modern Art simply had an exquisite Matisse exhibit that included a restaging of the well-known 1905 Paris Salon d’Automne, the place a critic dismissed Matisse and his mates as “wild beasts.” It jogged my memory of all of the “pictures?!?!?” drama 50 years earlier, additionally round one of many Paris Salons: The French poet Charles Baudelaire had the spiciest takes on photography, which entered the Paris Salon for the primary time in 1859. Considering pictures as artwork was, in his eyes, clearly a symptom of inventive decline. Counting a mechanical course of as artistic was not simply unsuitable! A risk to French style! Photography is “the refuge of each would-be painter, each painter too ill-endowed or too lazy to finish his research”.

Maybe I’m too ill-endowed in talent or too lazy to revamp my whole homepage from scratch. And but, right here I sit, being fairly pleased with my new homepage, made totally by means of mechanical means. I’m not French, and I’m unsure I’ve extra style than anybody else – however it’s precisely the best way I noticed it in my head, a mere 48 hours in the past, and I’m fairly excited to point out it to my mates.