1:40
DreW produced a 100-second animation where every frame was generated with JS and Python, using no video or image models. Multiple teams of agents worked around the clock for 19 hours, and the result is a demo of code-only animation nearing solved motion.362 saves on X
A 100-second animation made entirely in code with agentsArt, video & music@devteamdrew58831362
0:42
Anthropic researchers discovered prokaryotic vault proteins that assemble into complexes larger than human vaults, and engineered shapes to create even larger nanocapsules, working with Claude.210 saves on X
Prokaryotic vault complexes larger than human vaults found with ClaudeScience@mgdurrant51565210
0:53
rewind made a Silent Hill-inspired horror game in about four days. Opus 5.5 ran in auto mode editing files and launching Godot, while Nano Banana and Meshy T2.5 turned grouped prompts into about 100 textured props.194 saves on X
Horror game built in four days with Claude Opus 5.5 and GodotGames@rewind0222423194
0:46
Camper asked Claude Fable 5.5 for an anime stickman sword fight in one prompt. It rigged both characters, simulated capes at 120 Hz, composed an original 120 BPM score, synthesized 228 sound effects placed on exact frames, and fixed its own mistakes, at a cost of $15.71.97 saves on X
One-prompt anime stickman sword fightArt, video & music@camperaigg1601297
39:15
Dave Jones of EEVblog gave Claude Code a single prompt, and it reverse-engineered the 121GW multimeter firmware into C in under 24 hours. The resulting firmware compiles, flashes, and runs on the real hardware; source is on GitHub.69 saves on X
Claude Code reverse-engineered a multimeter firmware into C overnightRobots & gadgets@eevblog1822469
1 / 4
Novelist 灯野ユヅル tested PrunaAI's P-H3-Image in ComfyUI on an RTX A4000, generating their heroine Safia in different expressions, positions, and outfits from a single Stable Diffusion reference. The model preserved face, hair, horns, and uniform while changing pose in about 40 seconds.4.2K saves on X
Local image model keeps a novel character consistent across posesArt, video & music@CZTfTv7F0E325902.9K2394.2K
0:08
Physicist Casey Handmer used Opus 5.5 to vibe code a personal radar that tracks the user's position, heart rate, and respiration rate. Two short clips show it working, and Handmer is now looking for buried treasure.1.1K saves on X
A personal radar that tracks position, heart rate and respirationRobots & gadgets@CJHandmer2K1281.1K
1:37
Toby built an original playable game in 24 hours using Opus 5.5, Flux 3 Max, GPT Image, Tripo, Blender and Three.js.935 saves on X
Solo-built an original playable game in 24 hoursGames@jdyu4777231.4K101935
It sorted 19,411 email threads, found 3,456 journalists, identified 336 I’ve worked with, and called 30 “warm relationships”
PR agent sorts 19,411 emails and finds warm media contactsWork@PRcarly30513672
Paulista, a software engineer, gave Claude access to a home-control PWA and purchase history. It buys groceries, butcher and fortnightly expenses and suggests recipes, so they just receive deliveries at home.595 saves on X
Claude orders the groceries, butcher and household expensesEveryday life@paulistazinn7659595
0:37
Abhishek built a sailing simulator and let frontier models steer and trim sails, feeding them the full world state every 5 seconds. Argon won, Astra came second, Opus didn't finish, and a small Qwen model ran aground.550 saves on X
Frontier models race each other in a sailing simApps & websites@shekitup2.3K115550
Ibragim, a researcher at Nebius AI, compared Opus 5.5 alone and paired with DeepSeek V4.1-Flash-1017 on 12 ML coding tasks. Accuracy fell from 81% to 67% and cost per task rose from $1.20 to $1.29; telling Opus its tokens cost 20× more cut cost to $0.71.457 saves on X
Claude Opus plus a cheaper model raised cost and dropped accuracyScience@ibragim_bad46737457
we booked 101 law firms and real estate offices in 8 weeks without pitching a single service in message one
101 law firms and real estate offices booked in 8 weeks with AIWork@draprints2137423
1 / 3
Wharton professor Ethan Mollick used Opus 5.5 to build a site where you can explore the fastest routes into every US city, find fair meeting points, and see city watersheds.380 saves on X
Interactive maps of fastest routes into US citiesApps & websites@emollick86142380
1 / 2
Zondark tested Claude Duet's voice mode on an hour-long drive to a client meeting. While chatting, it acted as an AI manager, dispatching sub-agents to find potential customers, research a market, and set up fuel-price monitoring, all without interrupting the conversation.356 saves on X
Claude Duet managed three work tasks during a driveWork@ZondarkAI58424356
33:10
Michael Guo turned the Amazon Rufus-Air paper into a 33-minute Chinese tutorial with running scripts: the same training recipe lifts instruction following from 33.6 to 76.9, and each chapter runs on a laptop.338 saves on X
33-minute tutorial on the Rufus-Air post-training recipeLearning@Michaelzsguo29262338
Manga artist ぬこー様ちゃん used AI to turn a rough manga sketch into a 3D background, capturing a room with iPhone LiDAR and rendering it in their art style. The tool now works from rough sketches and can be opened in CLIP Studio Paint as a 3D layer or driven from Claude Code.322 saves on X
Manga backgrounds from a rough sketch, generated as 3D layersArt, video & music@nukosama52555322
1 / 2
Security researcher Nick Mykhailyshyn earned $19.5k in bug bounties this quarter by pairing agentic lead generation and validation with manual testing. Opus 5.5 built custom tooling, while Opus 5, Daybreak Blue, and DeepSeek V4.1 Flash 1017 filtered leads; the key bugs were found manually.285 saves on X
Agentic workflow earns $19.5k in bug bounties in one quarterScience@whoareme3341811285
1:15
观默, a former big-tech programmer and AI engineer, built a 75-second animation with no image, video, or audio assets. A single 170KB HTML file uses Canvas 2D and Web Audio to draw and synthesize every frame; Opus 5.5 architected the pipeline.258 saves on X
75-second animated short drawn and scored entirely in codeArt, video & music@guanmo_ai23216258
GIF
MIT CSAIL researchers report 'Weight-Space Programming': frontier LLMs like GPT-6 can read, fix, and rewrite the weights of small models such as Qwen3.5-0.8B, using interpretability tools or direct editing. The work suggests LLMs may one day optimize neural networks at human or super-human level.114 saves on X
GPT-6 edited neural network weights directlyScience@MIT_CSAIL34873114
0:24
David Freeman built After the Wind, a Ragnarok Online meets Monster Hunter MVP in Godot over five weeks, using Claude Code (38.5B tokens), Codex (2.9B), 112k tool calls and up to 8 agents at once.2 saves on X
After the Wind: a five-week Godot MVP with consistent visualsGames@bykaralord210242
1:24
Zhee built a Snapchat mini-game where you drive, refuel, pick up survivors and reach a safehouse before nightfall, using AI tools solo; it has 40,000+ views and 39,000+ plays.1 saves on X
A zombie survival game built solo with AI reached 40k Snapchat viewsGames@zhee_explores26691
0:06
Blake Shao packaged their motion graphics prompting workflow into an open-source skill called /blank-page. It expands a rough idea into expressive first drafts, like Midjourney for motion.1.9K saves on X
One skill turns rough ideas into motion graphic draftsApps & websites@blak3shao1.2K421.9K
Adam Holter had Codex analyze 120 generated videos and found none contained low-frequency airplane Doppler curves, suggesting a possible forensic marker for AI-generated footage.1.6K saves on X
Codex checks 120 AI videos for plane Doppler curvesScience@AdamHoltererer8K3521.6K
Exam candidate níkolas used Claude to create a structured notebook of their mistakes, turning errors into a study tool. The photo shows a personalized page; they say AI is optimizing their studies.1.5K saves on X
Claude builds an error notebook for exam studyLearning@_nicksuxx2.1K411.5K
1 / 4
Nyuriwo released a web app for blending art styles: select tags, adjust strength, and output prompts for different image generators, including PIXai and NAI.1.1K saves on X
Art style mixing web app for PIXai and NAIApps & websites@nyuriwo_ssan830791.1K
0:23
メーン会場 built an embroidery data creation app with Claude that converts images to stitch data and lets users specify satin direction intuitively. It replaces software that normally costs 100,000 yen.1.1K saves on X
Embroidery data app built with Claude replaces pricey sewing softwareApps & websites@ktzgw1.7K2421.1K
1:00
Kriks fed Claude Opus 5.5 a set of videos and had it analyze style, movement, lighting, colors and scene structure, then built a system that recreates the same style for completely different videos. The final result took two Ultra x20 subscriptions and a lot of work but was impressive to them.793 saves on X
A system to replicate a 3D cartoon animation styleArt, video & music@kristerssulckis64442793
0:37
Julian, an AI filmmaker and motion designer, built a Claude skill with Google Street View access to pull detailed reference photos of real locations for AI video work.426 saves on X
Claude skill scouts real locations via Google Street ViewApps & websites@sirlifehacker31519426
1 / 3
Malariadx tested the ComfyUI Image-to-3D model Idol on anime art, finding it works better with shaded images and also produces accompanying videos.424 saves on X
Idol turns anime images into 3D models in ComfyUIArt, video & music@malariadx33118424
0:15
Onofumi, developer of ZIVU AI, created a realistic full-body 3D Gaussian splat from four AI-generated images (front, side, back) in about two minutes, with no shooting or modeling.392 saves on X
Full-body 3DGS from four AI images in two minutesArt, video & music@onofumi_AI44840392
1:52
Game developer After The Wind shows their pipeline for creating monsters: a doodle becomes a Codex concept, Tripo P2 3D model and texture, Blender rigging with Claude, then into Godot.309 saves on X
Game monster pipeline from doodle to GodotGames@afterthewinddev28514309
0:29
ery_treasure used one line of Claude Code to generate a moving 3D model of a drum washing machine from a GitHub skill, with separate parts.251 saves on X
One-line 3D model of a washing machine in Claude CodeLearning@ery_treasure31829251
0:31
Gabriel Lattanzio, an Americanist professor at Sorbonne, woke early to build an interactive L3 course presentation on the US elections with Opus 5.5. A public version is available to explore.144 saves on X
Interactive US elections lecture built with Opus 5.5 by a Sorbonne professorLearning@GabLattanzio27228144
3:29
Kelly Boesch wrote the lyrics for "Only The Good Things" and produced the song with Suno, generated the imagery with Midjourney, and animated it with Veo.123 saves on X
A Suno song with Midjourney and Veo animationArt, video & music@kellyeld475104123
0:30
J A Z I I used Fable 5.5 to one-shot a tree animation in pure code, taking about an hour. The model generated the code for the animation.108 saves on X
Tree animation generated from a single code promptArt, video & music@notjazii38910108
GIF
Japanese illustrator hiyodori used ChatGPT to create a GIF of their character Chiharun dancing, finding it more intuitive than video generation.28 saves on X
Dancing character GIF generated with ChatGPTArt, video & music@hiyodorixxxx5126428
0:15
AI creator Zeeshi used Seedance 2.5 to create a cinematic fantasy film of a dragon flying over a medieval city, with a detailed prompt for physics, crowd reactions, and camera movement.8 saves on X
Fantasy film of a dragon flying over a medieval cityArt, video & music@AIwithZeeshi28138
1 / 2
Estagiário da F1 used ChatGPT to generate a speculative Audi livery for the Brazilian Grand Prix, referencing the team overalls, clothing line and Revolut card for Brazil. The post includes two mockups they will compare with the real reveal.7 saves on X
A ChatGPT prediction for Audi's São Paulo GP liveryArt, video & music@EstagiariodaF157487