Made by Mechanical Means

One noteworthy distinction in between dealing with Claude and the other performance tools I’ve touched is that function demands are remarkably not the # 1 conversation subject. Rather, individuals wish to know how I’m utilizing Claude, how they ought to be utilizing Claude, and whether there are any cool magic spells I can share.

In this post, I ‘d like to stroll you through the procedure of upgrading my homepage without touching or running code or perhaps utilizing conventional Claude Code. No PRs were made, nor did I open GitHub. If you have not seen the complete outcome, possibly inspect it out initially

Approximately 60 threads worked, typically in parallel, on my significantly outrageous demands, that included producing a CGI parody of a Werner Herzog documentary in Blender and manufacturing soundtracks from scratch. Regardless of the intricacy, a significant quantity of that orchestration occurred on my phone instead of my computer system -in reality, I didn’t need to run any code in your area. Claude worked completely in the cloud.

I’m composing this in late September 2026, and I think my workflow will look really various in 6 months. Even compared to the start of the year, I’m seeing a plain distinction – my work is now a lot more innovative, less bounded by the innovations I’ve mastered, therefore far more visual and psychological. I’ve utilized Opus 5.5 for all the work.

I began by developing a brand-new Claude Project. Jobs are ending up being a layer that covers Chat, Cowork, and Code; you can think about them as sitting above domain-specific ideas. You speak to one Claude with the capability to spin off extra sessions, share files and memory in between them, and monitor the whole fleet of Claudes to accomplish objectives.

Previously this year, I would have joked that I get a lot done by informing a lot of Claudes to do work. Now, I’ve put one Claude in charge of each of my tasks, and in turn, that Claude understands how to handle jobs, examinations, and threads. We’re talking more about objectives and less about how to attain them.

Every thread gets its own computer system in the cloud, where it sets up whatever it requires (Blender, FFmpeg, Playwright ), inspects its deal with screenshots from a headless web browser, and releases a personal sneak peek page for me to click through. All threads share one folder of files and one memory. Claude is clever enough to make a note of choices I made in memory and makes certain recently begun threads follow them without me needing to duplicate myself.

Conceptualizing a brand-new instructions

I began the method I ‘d begin with a style group: by disposing a lot of motivation on the table. Those consisted of 90s rave posters, films I liked, advertisements that spoke with me, images I keep returning to, even simply principles I like.

10 threads checked out those instructions in parallel with really little assistance. One developed portfolio instructions for someone who truly likes outdoor camping in national forests; another pictured what an 80s Ghost in the Shell site would appear like. Both, it ended up, felt a little exaggerated to me. I wished to make something brand-new, not simply clone something currently out there.

Whenever I liked an instructions, I asked Claude to double down, get imaginative, and offer me 5 more variations of the very same instructions.

While many instructions utilized HTML, Three.js, and other libraries to attain a few of my hopes and dreams, I quite rapidly discovered myself pressing Claude to utilize Blender. Claude does not click around Blender’s user interface, rather composing code and utilizing Blender as a Python module. When I quickly thought about turning my site into a relaxing camping site, Claude immediately designed one.

Offered the instructions of classic innovation I liked( Game Boys, record gamers, Walkmans, and so on) and the job of developing an innovative portfolio gallery, Claude ultimately recommended VHS tapes.

Together, we integrated Claude’s concept (VHS tapes) with my own vision for the site, that included an interactive variation of the sort of space a 90s German motion picture would provide a reporter.

Claude’s 2nd draft, including my instructions about the space setup and support to utilize Blender artistically, was shockingly excellent. The variation I delivered took approximately 60 versions, however the 2nd draft, which got here 39 minutes after my very first quick, is currently recognizably the space you see today.

Early on, Claude chose that the code ought to be modular, permitting threads to deal with aspects of the space 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"install( canvas, ctx, width, height, audio) return frame( t, dt) * draw one frame *, input( occasion) * play, stop, next, prev, click *, stop() ,;,;

After every model action, Claude released a brand-new variation of an artifact, a personal websites where I might see the present variation, both on my computer system and my phone. I continuously requested a couple of alternatives, clicked through them, returned some review, and happened with my day. Doing that from my phone felt especially effective. I would be on my bike, have some concepts while riding, rapidly fire off some concepts to Claude, and have a look at the outcomes on the next coffee stop.

Constructing the space

Let’s speak about the space, which is most likely where I pressed Claude to be most enthusiastic. Among the most lovely parts of LLM tools is that they permit me to do things I would not have actually had the ability to do before – I have no concept how to utilize Blender. I explain the important things that I desire, Claude composes the geometry as code, renders a still, and I provide feedback.

The space is now ~ 3,000 lines of Python code, managing Blender as a library. A few 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 reporter’s desk. A phone with a microcassette voice mail standing on a phonebook, a Rolodex, floppies, a pager, a solar calculator, a Discman, and a Walkman, with cables running behind the desk to an electrical outlet.
• retro_pc.py (431 lines): a late-80s beige PC clone from a fabricated brand name, with 2 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 space behind the cam, which you just see in reflections and in free-look: a four-panel door with a type in the lock and light under it, a Bakelite light switch, a stucco cornice, and a tiled range.

Claude chose its own when to search the Internet for easily certified designs and when to design props from scratch. Some furnishings originated from Poly Haven (CC0-licensed), however a lot of products are “made by Claude” – like the desk, the television & & VCR, Discman, Walkman, the whole old PC, or the whole space geometry. The surface area textures are code, too: Claude’s wood, plaster, plastic, and metal are procedural sound, not photos.

I kept promoting a photorealistic appearance, which quickly had us speaking about light maps. Real-time lighting in a web browser can’t do the soft bounce light I desired, so Claude chose to utilize Cycles, Blender’s course tracer, and bake the light into textures. It’s a remarkably included procedure: Every fixed surface area gets a 2nd UV set loaded into atlases, and each atlas gets up to 4 maps: its color without light, the light from the space’s own lights( desk light, city radiance, moon), the light from the television alone, radiant plain white, and the space with the desk light turned off. The shader tints the television’s light with whatever the tape reveals, so a blue tape turns the space blue:

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

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

Not to proclaim Claude’s horn here excessive- however it’s outrageous to me that all of this can merely take place in code.

For what it’s worth, while it’s quite cool to see Claude operate in the cloud while I’m riding my bike around, extreme 3D work actually gains from devoted GPUs. You can connect your own computer system in any thread by striking the little “+” icon and picking Work Locallywhich enables Claude to utilize your computer system for faster making.

The view outside

I had really particular concepts about what the “outside” need to appear like: photorealistic, alive, without consuming a lot of CPU/GPU cycles on visitors’ devices. You’ll be unsurprised to hear that as soon as again I didn’t actually need to understand what ‘d work best; we merely constructed each and every single alternative.

The very first placeholder was a picture (“Berlin Skyline” by Billie Grace Ward; CC BY 2.0). I believed I ‘d be smart and create a whole horizon in Blender. It looked respectable, simply not rather as excellent as the picture, so I deserted that instructions.

. Claude had a much better concept: We utilized Depth Anything V2an open depth-estimation design, and turned the image into a 50 KB depth map with 2 channels: a soft one, so masts, signals, and near structures slide versus the horizon when the electronic camera wanders, and a sharp one, so trains can vanish behind structures.

That offered us the very best of both worlds – the noticeable picture however likewise depth to make it possible for animated lights and trains. They’re all drawn by the GPU on top of the picture, at nearly no CPU expense:

Developing the nihilistic penguin

Let’s discuss my preferred part of the whole job: producing a parody of the “nihilistic penguin“a scene from Werner Herzog’s Encounters at the End of the World: one penguin leaves the nest and strolls inland, towards the mountains, alone.

Claude and I began easy, with a childish animation drawn completely in SVGs. It permitted us to talk about precisely what sort of parody I desired.

As soon as we had a respectable concept of what I desired, I asked Claude to develop a CGI variation that would initially imitate the initial carefully, so that we might include my modifications later on. I’m informed that Claude did what a VFX artist would do:

• It broke the scene down into 8 shots, cut on the original’s shot limits.
• It tracked the cam frame by frame and attempted to match every pan, zoom, and wobble of the portable video camera.
• It constructed (surprise, surprise) the world utilizing code in Blender, when again utilizing the “Blender via Python” method. This consists of whatever you see – snow, mountains, gravel, penguins.

It rendered 2,787 private 640 × 480 frames at 23.976 fps, which led to a 116-second long video.

Any excellent parody flourishes on being quite near to the initial with a couple of little information altered: a desk with a computer system on the ice, a row of business flags throughout the lonesome penguin’s method (one for each location I’ve worked), or a ridiculous recommendation to the reality that I have a blog site that gets a brand-new post like when a year.

While Claude was total outstanding at sound style, it at first attempted some complimentary open-source designs to produce a voice over. It worked however was a little lackluster in regards to quality, so I registered for an ElevenLabs Starter account and just provided Claude the API secret.

Manufacturing music

Oh, likewise, did I discuss that Claude made up and manufactured 3 soundtracks on the site from scratch? It utilized (once again, surprise!) Python. It initially composed a rating, then a synthesizer engine. I have no concept how to make music. I asked Claude about its method: “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 song form.” I have no concept what any of that suggests however I had the ability to offer Claude Rick Rubin-style feedback in the kind of “I like this” or “No, make it… less excited”

Not checking out the code

I believe a great deal of engineers are experiencing a crisis of faith: If I’ve got lots of Claude threads buzzing away at the very same time, there’s no other way I’m checking out all that code. Which holds true: I’m not checking out the code. I think that this is much easier to comprehend for individuals who’ve worked both as an engineer and as an engineering supervisor – as a supervisor, you need to discover methods to develop quality without examining each and every single modification your individuals make. Whole books have actually been discussed how precisely you do that without micromanagement, however I use approximately the very same systems. 2 timeless examples are requiring extensive procedures and measurements, or charging private representatives with breaking things (aka adversarial representatives who consider their success to show that the code is bad).

Numerous QA threads drove the website round after round in a headless web browser and discussed 3,000 lines of findings to the shared folder. The collaborating Claude rapidly determined which other thread was at fault and passed the finding on. The multi-Claude world is, as it ends up, not blameless. The adversarial representatives discovered cassettes with blank labels, remaining German, links that remained clickable behind cards, and a race when you change tapes rapidly. A single Claude thread does not compose ideal code, a fleet of Claudes holding each other responsible fulfill my bar.

Hill climbing up efficiency

Designs are extremely proficient at hillclimbing. Entrusting threads with particular targets, like frame rate, draw calls, or shadow draws, gets you extremely far.

You ‘d believe that designs have a tough time informing whether 2 variations of a scene look approximately the very same. Claude developed a smart option – it took relative screenshots and in fact determined the pixel-by-pixel distinction. If you do that from several angles, you get a relatively sure-fire technique of making sure that the efficiency gains are genuine gains and not simply quality deteriorations.

Some closing ideas

I wish to be delicate to those who may be reading this post with a pains in their heart, seeming like it’s outrageous that someone who never ever went through the discomfort of finding out the theory or practice of 3D modeling all of a sudden declares they’re developing 3D things. I’m sure you might discover single aspects that are poor quality and call them slop, mentioning distinctions in between the important things I dealt with for 24 hours and a space developed fastidiously by hand over weeks.

I’ve definitely heard comparable sensations from engineers, both when Electron made desktop app advancement much easier and clearly throughout the increase of AI-assisted coding. I’m most likely a relatively average engineer, however I am an expert one – I’ve dealt with lots of things that made it to lots of users. In 2026, I would not have actually thought about restoring my site without AI, however it’s amusing that I didn’t even grab Claude Code as a devoted “for coding” tool. I simply utilized Claude. Opus 5.5 and the generic harness didn’t require to be informed that we’re “coding” here.

The San Francisco Museum of Modern Art simply had a fantastic Matisse show that consisted of a restaging of the popular 1905 Paris Salon d’Automne, where a critic dismissed Matisse and his good friends as “wild beasts.” It advised me of all the “photography?!?!?” drama 50 years previously, likewise around among the Paris Salons: The French poet Charles Baudelaire had the spiciest handles photographywhich got in the Paris Salon for the very first time in 1859. Thinking about pictures as art was, in his eyes, undoubtedly a sign of creative decrease. Counting a mechanical procedure as innovative was not simply incorrect! A hazard to French taste! Photography is “the refuge of every would-be painter, every painter too ill-endowed or too lazy to complete his studies”

Perhaps I am too ill-endowed in ability or too lazy to revamp my whole homepage from scratch. And yet, here I sit, being quite pleased with my brand-new homepage, made totally through mechanical methods. I’m not French, and I’m uncertain I have more taste than anybody else – however it’s precisely the method I saw it in my head, a simple 48 hours back, and I’m quite thrilled to reveal it to my buddies.