LINKEDIN VERSION
Build a 3D explainer people can take apart, with Claude Opus 5.5
Ryan Sael (@RyanSael) posted a lens lab on September 23, a 3D lens you play with in the browser: move the ring and watch the plane of focus move. It has 3.4M views, and the Claude Opus 5.5 run behind it cost him $25.66. After reading this, you'll be able to build one for a machine you pick: a 3D model people can cut open, operate and take apart, with a readout that explains what they see. I built a pin tumbler lock this way from an empty folder with five plain messages to Claude Opus 5.5, and it took 74 minutes from the build request to a finished page.
[Video 01: 01-pin-tumbler-demo.mp4] My run, 14.7 seconds: the wrong key stops, the right key turns, then the lock comes apart
That's my lock. The wrong key stops because pin 3 sits 0.045 in too low and pin 5 the same distance too high, and the page says so.
Here's Ryan's post, with the bill in the caption.
[Image 02: 02-post-2102591147927654847.png] Post by @RyanSael on X
Konstantin (@konstantinsaifo) did the same for a Raptor 3 rocket engine and reached 853K views, then open-sourced the lab the next day.
[Image 03: 03-post-2104094723887501736.png] Post by @konstantinsaifo on X
The captions make it sound like one sentence, some version of "I asked Opus 5.5 to explain [the mechanism] by building an interactive [object]." But Ryan's prompts also ask Claude to scan his files and unpublished projects, and he said that with zero context it "probably won't achieve the same quality." Sourany Phomhome's published one-shot prompt for a koi pond scene runs about 16,600 characters.
What you need
- Claude Code or Cursor with Claude Opus 5.5. Victor Mustar (@victormustar) set effort to xhigh for his LEGO Microduck; I used Cursor, which has no effort setting.
- A paid plan for that tool. Konstantin built his Raptor on Claude Pro at "around 20% usage," and Ryan's usage bills ran from $25.66 to $78.19.
- About an hour and a half, mostly waiting.
- One machine with published numbers, like a manufacturer's chart. I used the Schlage Classic key chart.
Here are the five messages on one card.
[Image 04: 04-cheatsheet.png] Cheat sheet: the five messages to send the agent, from describing the page to checking its numbers
1. Tell the agent what you're making
Start in an empty folder. Parts built in code, instead of a downloaded model, get names the page can point at later, like "pin 3". Let the agent run commands without asking each time, or you'll come back to a session stuck on a permission prompt.
I want to build an interactive 3D page in the browser that explains how a five-pin pin tumbler lock works. Set up an empty web project for it with three.js, and build every part in code, with no downloaded models. Make sure you can run and screenshot the page yourself, because you'll be checking your own work.
You get back an empty page that runs, and the agent can screenshot it.
2. Have it write down the numbers and the rules
Claude Opus 5.5 writes a program, so the page works out why the wrong key stops instead of playing back a clip. That only matters if the numbers are right, so the habit this method rests on is keeping every number in one file, next to its source. Then anyone can check a readout, and that's what a company pays for.
Before you build anything, find the published numbers for this lock: where each cut sits on the key, how deep each cut can go, and how long each pin is. Write them in one sources file with the source next to every number, and list anything you can't find separately as an estimate. Then write the rules for this project into CLAUDE.md: every number comes from the sources file, every part is built in code, and the job isn't done until you've screenshotted every state and written down the three worst problems you can see.
Claude Code reads CLAUDE.md, a plain notes file, at the start of every session in that folder. My sources file came back with the Schlage chart, a source beside every number, and a list of unknowns the page must label as estimates, from spring force to how far a blocked plug turns before it binds.
3. Send the build request and walk away
Every message here is shortened from what I ran. My real build brief was a full page, with eight named screenshots and rules for how the metal should look.
Read CLAUDE.md and the sources file, then build the lock as a 3D page. The control that explains it is the key: no key, the right key or a wrong key. With the right key a Turn control rotates the plug, and with the wrong key the plug stops and the page says which pins block and why, using numbers the simulation works out. Add a cutaway slider, a take-apart button, labels, and a card for any part I click with its name, what it does and one number marked measured or estimate. Make it look like real brass and nickel, not a clay render. Work alone for at least an hour and stop by two: screenshot every state, look at the pictures, write down the three worst problems, fix them, and don't ask me questions.
Swap in your machine and keep the rest, but don't point the agent at Ryan's page, which he hasn't licensed. My first screenshots landed 21 minutes in, and the agent kept fixing until it stopped at 58. If yours stops early with a summary and no screenshots, tell it to finish them.
[Image 05: 05-wrong-key.png] The wrong key in my lock: pins 3 and 5 marked as blocking, the readout listing each pin, and the plug stopped at 1.5 degrees
4. Fix the three worst things
The agent's own list of problems becomes your next message, capped at three so it doesn't restyle the whole page.
Look at the screenshots again and fix only the three worst problems you wrote down. Don't restyle the page and don't add parts. Take the screenshots again and tell me what's still wrong.
Mine named a key cropped at the edge of the frame, a cut face that looked like wood grain instead of brass, and a dark, mottled plug interior. All three were fixed in under two minutes.
[Image 06: 06-turned.png] The right key turned 90 degrees after the fix, with the cam turned along with it
With a night to spare, try what Vox (@Voxyz_ai) did with his black hole lab: three review rounds while he slept, 5 hours 28 minutes and $90. He didn't share the prompt, but he described the loop: "Two reviewers each picked 9 to 10 issues and ranked them P0, P1, P2. The third one fixed them." It also backed up before each round and rolled back anything it couldn't fix.
[Image 07: 07-post-2103117246860345550.png] Post by @Voxyz_ai on X
5. Try to prove the numbers wrong
A page can look perfect and still be wrong. Victor's LEGO Microduck promised "0 collisions, every step buildable" and reached 1.3M views.
[Image 08: 08-post-2103110908444631120.png] Post by @victormustar on X
Then Tim Krell (@timothykrell) found a step that flips a part "as if it were a contiguous piece, but there's floating parts," and Victor replied "your findings seem legit". Have the agent try to break yours before a stranger does.
Don't change the picture. Check the numbers instead. Work out by hand from the published chart where each pin should sit with the right key and the wrong key, then set the page to both keys and put its readout next to your arithmetic. Drag the cutaway through its whole range and make sure nothing breaks. Then mark every number in the help text as measured, estimate or unsourced, and fix anything unsourced.
My run did a looser version with three pins by hand, and it passed: the arithmetic, the on-screen readout and the 3D parts all agreed. It also caught the two sources disagreeing on which end of the key is pin 1, so the page follows Schlage's own sheet and says so. When a client's sources disagree, pick one and say which.
[Image 09: 09-close-wrong.png] Close-up of the wrong key: driver pin 3 still crosses the shear line, and key pin 5 is pushed past it
When the check is clean, ask for the final page as one file that opens offline, drag every control yourself, and put it online. Ryan simply asked Claude to put his on Cloudflare Pages.
What I got
I ran all of this in Cursor. It took 74 minutes from the build request to a page that opens offline: 58 for the build and about 16 for everything after it. Cursor showed me no usage figure, so I can't tell you what the run cost.
Before I sent a single fix, the agent had already caught an exploded view that didn't fit the frame, overlapping labels and mottled metal.
[Image 10: 10-exploded.png] The lock taken apart: key, housing, plug, pins, springs, cover, clip and cam, where the clip is a thin arc and the cam a plain disc
Plenty is still wrong. From outside, the lock is a plain gray block, a few label letters peek out past the open help panel, and the key flashes white about 11.3 seconds into the demo. It held 120 fps on an M4 Max, but I haven't tested a phone or Safari. The whole page is one 793 KB file.
I wouldn't post it next to the lens lab, though. Ryan's has a bench, glass and a lit room, and mine is a lock in an empty dark studio, so I'd send it to a client as a lesson. Closing that gap would take a set and a camera move, probably the kind of extra turns that took Ryan's museum from $29.78 to $78.19.
Making money from it
Nobody in these threads posted a client invoice, so there's no going rate yet. The nearest published prices are Hon Tran's $4,000 to $10,000 for a 3D configurator of one product, assets extra, and Utsubo's $20,000 to $40,000 for a hero scene or around $20,000 to $50,000 for one museum station. My suggestion is to pitch one machine and one question to a hardware launch, a furniture retailer or a museum, for a fixed fee of around $4,000 to $10,000. Konstantin gave his code away under MIT, so what you sell is the checking: every number tied to a source the client approves, plus his README's warning that these are teaching models and shouldn't be treated as engineering data. I'd put Ryan's line in the statement of work: "people trust what looks good. so what looks good had better be true."
Credits
The creators: @RyanSael for the lens lab, @konstantinsaifo for the Raptor and airsup-lab, @victormustar for the Microduck, @timothykrell for testing it, @Voxyz_ai for the overnight loop and blackhole.voxyz.ai, and Sourany Phomhome for the koi pond prompt.
Posts quoted above: Ryan on zero context, scanning his files, the museum bill, trust and Cloudflare Pages. Konstantin on Claude Pro usage, going open source and in his README. Victor on xhigh and Tim's findings.
Docs and data: Claude Code's memory page, and the Schlage Classic chart in Tyler J. Thomas's Key Bitting Specifications and Schlage's key specs sheet.
Prices: Hon Tran, Utsubo on agencies and museum installations.
View counts as of September 28.