2026-09-24
Ten things AI built while I slept — the difference wasn't the prompt
A Threads post listed 20 things people built with Claude Opus 5.5. Overnight, I had ten of them rebuilt around my own site, lectures, and data. You can run every one of them on this page.
On September 24, 2026, a Threads post collected 20 things people had built with Claude Opus 5.5 — games, 3D worlds, pixel animation, even music, all from text instructions. The most common reply wasn't awe. It was "why doesn't this work when I try it?"
I had the same question. So before bed, I handed the thread to Claude Code with one sentence: "Pick the five cases most relevant to our website, to me, and to my work, build one result for each, and write it up as a report I can read in the morning." After reading the report, I asked for five more.
All ten results are below, and every one runs on this page. The short version: the prompt wasn't what separated good results from bad ones. The materials the agent had access to were — and the habit of looking at the output and fixing it.
The demos keep their original Korean interface. They were built for my Korean-speaking audience, and I've left them as they shipped.
One sentence was the whole instruction
The agent read all 20 cases first and set its own criteria. Could the result go straight onto my site or channel? Could it be filled with my real data? Did it avoid repeating a technique already used? Games built on other companies' IP — Minecraft, Roblox, Mario — didn't fit the brand and were dropped. So was the Unreal Engine case, since Unreal isn't installed on this machine.
It then started five sub-agents at once. Each one built a single piece, rendered it in a browser, looked at the screenshots, fixed what it saw, and reported back. Everything was committed to a work branch; the live site wasn't touched. The second five followed the same process.
The ten builds
01. A launch video for this site
The original case produced a one-minute product launch video in a single request with the HyperFrames skill. Mine captures careerhackeralex.com itself — lectures, Second Brain, prompt generator, 3D studio — in 51.5 seconds. Every number on screen is one the site actually shows.
02. How this website grew
The original had Claude animate its own history, drawing every frame in JavaScript. Here the website tells its own history instead. It stacks all 815 commits, day by day, from the first on April 21, 2026 to September 23. The final scene uses the co-author lines on 595 of them to show which Claude models I worked with, and when. The sound is synthesized in code too.
03. Freedom run
The original was a pixel runner in space, built with Canvas alone and no external assets. Mine turns my brand line — "freedom from time, money, and emotion" — into game rules. Dodge the late-night clock, the credit card bill, and the anxiety cloud. Pick up the lecture, the prompt generator, and the Second Brain to refill your gauges. Every sprite is a pixel array in the code.
04. A portrait painted in code
The original repainted travel photos as acrylic paintings using JavaScript, with no image model. I gave it my profile photo, and it painted three plates: charcoal hatching, a two-ink riso print, and a warm-grey grisaille. The agent looked at its own output and revised the algorithm seven times. That's how it caught a version where my dark hair vanished into the background and I looked bald.
05. The AI-native island
The originals were Three.js games — an open world and an underwater exploration. Mine is a 3D island with one landmark for each of the 12 chapters of my AI-native course. Walk into a landmark and its chapter card opens. The chapter list comes straight from the course data.
06. The clay desk
The original drove Blender to make a claymation. This 8-second loop comes from a single Python script that handles modeling, materials, lighting, animation, and rendering. A developer hands a late-night clock, a bill, and an anxiety cloud to his laptop agent, one at a time, then leans back.
07. Monday, 9 a.m.
The original wrote its own story, drew its own frames, composed its own music, then reviewed the film second by second. Mine is a 60-second short for my lecture opener. Every frame is drawn in JavaScript; the music and sound effects are synthesized in Python. The agent reviewed a one-second contact sheet and loudness measurements, rendering the film three times. All seven scene cuts land within 7 milliseconds of a musical note.
08. The prompt sketchbook
The original turned typed descriptions into drawings you could move with the keyboard. Mine is a teaching toy. It opens with "a cat" and "a big cat in a red scarf" side by side, so the lesson is visible before anyone types: the more specific you are, the more the result changes. On this page it draws from 25 built-in doodles and understands Korean modifiers.
09. Draw a workflow, watch it roll
The original turned a hand-drawn sketch into a 3D simulation with physics. In mine, ramps you draw on paper stand up as 3D walls. Balls labeled email, report, and meeting notes roll down them into one of two bins: "automated" or "needs a human." It's a toy for feeling, with your hands, that automation is a design problem — not a tool problem.
10. A 1996 office, a 2026 desk
The original researched maps and photos of San Francisco before the 1906 earthquake, then rebuilt the street in 3D. Mine researched Korean offices of the mid-1990s first — 14 claims, each graded by how well it is sourced — and only then built two desks. A slider moves between the eras, so you can see what the paper approval folder turned into. Every card carries its source numbers and a confidence grade.
Three materials made the difference
The first was real data. 02 came from git history, 05 from course data, 04 from my photo, 01 from the live site. "Make a cool animation" doesn't produce any of this. My materials were already somewhere the agent could read them — and that's why the results are mine rather than someone else's demo.
The second was tools already in place. The video used the HyperFrames skill, the clay used the Blender installed on this machine, and the 3D used Three.js. The agent didn't invent tools. It picked the ones that were there.
The third was looking and fixing. Given my one sentence, Claude Code wrote its own brief for the sub-agents, and that brief required each one to render its work, look at it, and fix problems before reporting. The portrait's algorithm was revised seven times; the short film was rendered three times. In 09, the agent traced ball paths to find a physics-engine setting that let balls pass straight through walls.
Judgment stayed with me
The agents were candid about their limits. The music in 07 was checked only by loudness numbers — no one has listened to it. Of the 14 research claims in 10, only 4 were confirmed by articles from the period; the cards mark the rest as inferred from common practice. The agent building 06 finished its work, then stalled just before its final report.
Choosing what to build, deciding what fits my brand, listening to the sound, checking the photos — that was still my job. The last card of 07 sums up the whole night: the work didn't disappear. It changed shape.
Hand off one thing tonight
If you want results like these, prepare three things before you polish a prompt. Put your materials in a folder the agent can read — slides, data, photos are enough. Add one line to your instruction: look at the result, fix it, then report. That line turns the agent's work into verified work. Then, in the morning, you decide what to keep.
All ten builds here were made in a single night. Which ones end up in my lectures and on my channel is what I'm deciding now.
Original thread: @choi.openai, "20 things built with Claude Opus 5.5" (Threads · 2026-09-24)