TALUS Virtual Store hangs 12 hats on a pegboard, then a rail of jackets.
Viktor used Fable 5.2 to build it in the browser. The clip moves through Caps, Bags, Shoes and Clothing, then opens a Glacier Quarter-Zip Fleece at $129, size M, and taps Add to bag.
I did not build this.
My view: a rack in the browser only works if you can still check out from the hanger.
(Video via @viktoroddy)
1
24
Seven spells, about 1800 physics pieces, and zero 3D models or textures.
Blendi used Fable 5.2 to make Small Magic in one day, and it runs in the browser. Everything in the clip is code: a meadow, a cottage, then debris from the spells. Seven numbered icons sit on the bar. The labels I read on screen are Explosion, Singularity, Tornado, Fire, Frost and Growth.
He says it is open source and playable now.
My view: the constraint is the hook. A scene made only of code still has to be fun to knock over.
(Video via @BlendiByl)
1
63
Play it here: patch.gg/play/blendi/small-m…
I could not find a public GitHub repo from the play page or from his posts. The game lives on Patch.
He posted earlier today that he had added an 8th spell, Time Rewind. The video on the root is the original clip with 7 spells.
8
Same prompt: "make a jelly orange slice." The prices on screen in his video are $7.57 for Opus 5.5 and $11.57 for Fable 5.2. I did not rerun either model.
He asked which you'd pick if the names and prices were hidden. Both sides show a playable "Citrus Jelly" toy: grab, stretch, let go, with sliders for firmness and damping.
In the clip, Opus goes glossier and slumps more. Fable keeps a cleaner wedge. My view: I would pick Fable with the labels hidden, because it still looks like a jelly slice, not a puddle. That costs more in this test. One clip is not a leaderboard.
(Video via @vib3coded)
1
1
72
I did not rerun either model. The $7.57 and $11.57 figures are the on-screen labels in his video, under Claude Opus 5.5 and Claude Fable 5.2.
Both sides show the same kind of toy: a "Citrus Jelly" wordmark, fruit-type buttons, and sliders for firmness and internal damping. I do not have a public demo link from the thread.
1
20
xeet rendered 241 X posts to PNG in 260 ms, with no browser in the loop.
That's Guillermo Rauch's new open source renderer, in Rust and WebAssembly. The demo is a Tetris game where every frame is a finished post image. On screen, single frames clock in at 1.0 to 1.4 ms.
He says zero browser, network, filesystem and system fonts. Apache 2.0.
My view: if you need a post preview for an agent or a CLI, this is the first thing I would try before booting Chrome for a screenshot.
(Video via @rauchg)
1
1
80
Demo: xeet-rs.vercel.app
Repo: github.com/rauchg/xeet
The README's own p50 for a text post at 390 px and 3x is 2.8 ms, or 4.5 ms with four media placeholders. The 1.0 to 1.4 ms figures in the video are per-frame times on that demo.
It ships a native CLI and a WebAssembly core. I would start in the playground before installing anything.
12
Shubham Saboo swapped his TV's ad-filled home screen for his own photo, a clock and four app names. One markdown file and a coding agent did it.
He ran it with Gemini 4 Carbon in Antigravity. The playbook is open source and also works with Claude Code, Codex, Gemini CLI or Cursor.
What the agent does over ADB:
1. switches off ads, telemetry and unused preinstalled apps, and uninstalls nothing
2. crops your photos so text never sits on a face, with face detection running locally
3. turns debugging back off when it finishes
Catch: Android TV and Google TV only. Samsung, LG and Roku won't work, and the README says he tested it on one TCL set.
Of all the agent demos I saw this weekend, this is the one I'd run tonight.
(Photo via @Saboo_Shubham_)
2
2
121
Playbook: github.com/Shubhamsaboo/awes…
You need a computer on the same Wi-Fi, a folder of photos and about an hour with the remote nearby.
Every change can be undone: Settings > Apps > See all apps > Show system apps > Enable. Menus differ by brand, so stay in the room while it runs.
1
14
Claude rebuilt Starship's tower catch in three.js, then said which parts it guessed.
Ethan Mollick's prompt asked Opus 5.5 to find the catch footage, rebuild the last 40 seconds, add cut-away blueprint views and sync it to the webcast. A Fable 5.2 pass polished it after.
When he asked for sources, Claude named four:
1. SpaceX's webcast replay, seconds 74 to 116, for timing and camera shots
2. Wikipedia for the ship layout in the blueprint view
3. Guinness World Records for the 146 m tower height
4. Community cut-away drawings for which way the tank domes bulge
It said everything else was estimated, including heights along the hull, tank sizes, pipe routing and catch rail height. No SpaceX engineering drawings.
My view: A 3D rebuild with no source list shouldn't be trusted.
(Video via @emollick)
1
2
130
Play it in the browser: ship-42-catch.netlify.app
You can switch to the blueprint cut-away or fly the camera around the tower yourself.
A tip from his replies, from Felix Rieseberg: for a closer match, have the agent measure pixel error between the video and the render and push it under 10 in 255.
22
A flat page from a grade 6 science textbook, and then the diagrams stand up off the paper and run the lesson.
That's Gemini 4 Hydrogen, in a clip from an educational technology PhD in Oman. A coffee filter, a sand and gravel column and a small pond rise as 3D models, walk through separating insoluble substances, and settle back into the flat drawings on the page.
His prompt's hard rule: keep every word, invent nothing. My view: that rule is why I'd hand this to a parent or a teacher. The page they already trust stays exactly as printed, and the model only adds the motion.
(Video via @smalkalbani)
1
62
A condensed translation of the prompt in his first reply:
Task: turn a real textbook page into an interactive 3D educational scene. Use the attached image as an original textbook page, and make a cinematic 16:9 teaching video in which the page itself is a realistic physical 3D object, with the educational elements rising from it and moving above its surface.
Keep the real page in frame, slightly tilted, with paper thickness, texture and shadows, and do not change any text, number, drawing, color or layout. The camera should slowly approach, the main diagram should lift from a flat drawing into a 3D model while staying visually tied to the page, and all motion should happen above the paper, with shadows falling on it. After the lesson, the models should ease back into their original flat drawings. Do not replace or redesign the page, do not invent new facts, do not put the video inside a screen or a rectangular frame above the book, and do not add logos, writing or music. He notes the curriculum as grade 6 science, for a student about 12.
(Video from @smalkalbani)
37
On Alibaba's own benchmark, same model, same 200 PRs: about 84% of Claude Code's review comments weren't real defects. For Open Code Review it was about 62%.
Open Code Review used 352K tokens per review, versus 2,062K. Average time: 1m6s versus 5m38s.
Both runs put Claude 4.8 Opus underneath. Recall: 11.7% vs 12.7%, so it misses a little more.
The trick is boring. Plain code picks the files, bundles related ones and matches the rules. The model only reviews.
Apache 2.0, one npm install, any OpenAI or Anthropic compatible endpoint. Alibaba says it ran internally for two years.
My view: I'll take fewer comments that are right over a wall of maybe-bugs.
(Chart from the open-code-review README)
1
36
Repo: github.com/alibaba/open-code…
The benchmark is 50 popular open-source repos, 200 real pull requests and 10 languages, cross-validated by 80+ senior engineers (1,505 annotated ground-truth issues).
Two details worth knowing:
1. If you don't want another API key, its delegate-review command runs on the model your Claude Code already uses.
2. The scan command reviews whole files, handy for a codebase you just inherited with no useful diff.
The README credits the "~1/9 of the tokens" average to its own benchmark. The 352K vs 2,062K pair above is the two Claude 4.8 Opus rows of that table.
27
Fable 5.2 recreated a three.js game trailer in Blender in a few hours, @MengTo says: a tabletop dungeon diorama, a torch-lit hall, an armored party walking through, and a flaming dragon emblem at the end.
His setup: Fable 5.2 (medium) orchestrating Opus 5.5 agents on high, which he calls his main workflow now. Since finding RunPod, he says, he recreates his 3D scenes in way more detail without clogging his MacBook.
He asked for Blender, not a video model.
My take: if you get the Blender file too, this beats a prompt-to-clip, because you can open the scene, relight it and render again.
(Video via @MengTo)
3
1
154
The setup, as he describes it: Fable 5.2 on medium orchestrating, Opus 5.5 agents doing the work, RunPod instead of his MacBook.
He started from the trailer of his own three.js game, the one he built with Opus 5.5 so an AI could run D&D for his kids, so the agents had a finished cut to match. That reference is the part I'd copy first.
40
A Souls-like browser game in 3 days, built by Fable 5.2 directing up to 9 Opus 5.5 agents in Three.js.
Meng To says a playable level and a boss took 6 hours and 12 prompts. Then the quality loop: every change got same-camera 4K before and after shots, a fresh judge saw them shuffled, and only wins or ties were kept.
He's open about the gaps. The video is a scripted bot, not him playing. Strict judges still score most of it 5 to 6 out of 10, it takes over 10 seconds to load, and the walk was rejected 4 times.
My view: the judge loop is the part worth copying first, because more agents only make more changes, and the loop is what decides which ones stay.
(Video via @MengTo)
1
1
240
The 6 steps, as he laid them out in the thread:
1. One prompt to Fable 5.2, with the dream-loop skill and his workflow-score-to-target and design-action-combat skills. 6 hours and 12 prompts later: a playable level and a boss.
2. Fable 5.2 coordinates Opus 5.5 side threads, one per area: lighting, effects, characters, props, UI, sound. Up to nine at once, each owning its files. He ran this in Claude Code. For sound he used his Seerveil library plus the earshot skill.
3. Concept art was turned into 29 Tripo 3D models. They came unrigged, so the agent wrote the rig and animations in code.
4. He gave the agents a target image to chase, told to keep exactly the same camera, framing, layout and pose.
5. UI from a few short prompts: "Give me a UI concept for the bottom bars as well, for the entire UI of the game". "Build it". "use more border gradients/uneven/shadowy/noisy". "generate transparent pngs for these icons".
6. Every change gets a 4K before and after from the same camera. A fresh judge sees them shuffled as X and Y, and only wins or ties are kept. Full rounds use two judges and a rubric out of 10. The lower score counts. The target is 9.
(Video from @MengTo's thread)
43


