Best AI Creations This Week: A Shooter From One Prompt (Nov 13)

The best AI creations this week (November 7 to November 13, 2026) were games and objects, not chat answers. Claude Fable 5.2 came out on November 9, and within a day Felix Rieseberg posted Saltline, a browser shooter he says the model built from one short prompt in eight hours, models, sounds and textures included. Lex Toumbourou turned his nephew's game idea into a 12-stage, four-player beat 'em up overnight. Anthropic showed three working gadgets whose circuit boards Claude designed, and Pietro Miotti sent 48 AI agents through lidar scans of the Maya jungle and found a hilltop site that turned out to be already published, which is its own lesson. I build with AI every week as a product designer and design engineer, so I picked for craft over hype, and every entry links its original post.

The best AI creations this week

Eight picks, each checked against the original source. Numbers appear only where the source shows them, and I say where they come from.

1. Saltline: a first-person shooter from one prompt

Felix Rieseberg, an engineer at Anthropic, posted Saltline on X on November 9, the day Fable 5.2 launched. He says he wanted to test the model with one-shot prompts and no human feedback, that everything in the game (models, sounds, textures) was made from scratch, and that the run took about eight hours and roughly $550 at API pricing. In a reply he shared the prompt, which is two casual sentences asking for "insanely detailed quality" in a tiny level. The game is playable in the browser as a Claude artifact, and the post showed more than 4 million views when I checked.

The part that matters for designers: the prompt was not a spec. The model filled in the art direction itself. That is impressive and also the warning. If you do not bring the taste, you get the model's taste.

2. Aiden's Revenge: a nephew's game idea, built overnight

Lex Toumbourou wrote up Aiden's Revenge on November 10, a birthday present for his nephew. It is a four-player beat 'em up with 12 stages, 12 bosses and 18 playable characters, including family members, the dog and the pet turtles. An earlier version made with Claude Code in June was, in his words, not worth more than a couple of minutes of play. This time he gave Claude Fable 5.2 one large prompt before bed, and it rewrote most of the code and produced the sprites, cutscenes and ending. Images came from ChatGPT Images 2.5 and the voice lines and sound effects from ElevenLabs. He is honest about the rest: features like a Hulk transformation came from follow-up prompts after playtesting, and he never tested whether Opus 5.5 could have done the same. You can play Aiden's Revenge in the browser. The Hacker News thread passed 500 points.

What I love here is the brief. A kid numbered the bosses. That is a better design document than most product specs, because it is specific about what is fun.

3. Claude designed three working gadgets

Anthropic published a research post on November 9 about three prototypes Claude designed from short prompts: a USB-C speakerphone puck, a home voice-assistant hub and a battery-powered handheld virtual pet. Working in Diode's design software and KiCad, Claude wrote the specs, drew schematics, placed parts, routed four-layer boards, designed 3D-printable enclosures, wrote firmware and made the parts list. When KiCad's router was too manual for the hub's long tracks, it wrote its own router. Diode engineers manufactured and assembled the boards, and Anthropic says all three worked without electrical modifications. The caveats are in the post too: these are early prototypes with no certification or thermal testing, and Anthropic has a commercial relationship with Diode.

My favourite detail is industrial design, not electronics: a wedge-shaped hub enclosure shortened the microphone ports so their resonance moved away from speech. Form solving a sound problem. That is product design.

4. 48 AI agents and a Maya site that was already on the map

Pietro Miotti wrote about pointing AI agents at public lidar scans of about 1,200 square kilometres of Maya jungle. He says 48 agents reviewed the shaded relief images looking for plazas, platforms and pyramids, with no special classifier, then searched archives and drafted a report that he sent to Mexico's INAH, which replied that no registered site existed there. The write-up hit the top of Hacker News on November 11 with over 1,000 points. Then came the honest update: the site had already been mapped in a 2025 paper the agents missed. A rediscovery, not a discovery. He is clear he is not an archaeologist.

I am including it because of that update. Agents are fast at looking. They are still bad at knowing what they have not read. Any research tool you design for them needs a "who already found this?" step.

5. Tiny València: a whole city repainted as a SimCity map

Tiny València, by the maker behind the X handle @13ug1mb (leugim on Hacker News), is an isometric, SimCity-style map of the entire city where every building is real. The site explains the method: real 3D city data is rendered from a fixed isometric camera, then an image model trained on a hand-picked pixel-art style repaints it tile by tile, each tile continuing its neighbours so there are no visible joins. Data comes from the Spanish cadastre, IGN, OpenStreetMap and Google's 3D tiles. The Show HN post went up on November 10 and passed 250 points. Commenters spotted some hallucinated details, and the maker replied that they had fixed some but not all.

This is the pattern I trust most: hard geometry from real data, AI only for the paint. The model never decides where a building is.

6. ZigZag: embroidery software with an AI agent on the canvas

Kath Korevec's $300 embroidery machine wanted a subscription, so she built ZigZag, free and open-source embroidery software for Mac and Windows, and posted it as a Show HN on November 12, where it passed 540 points. She had been sharing progress on X since October. It turns SVG art into stitches, lets you edit them like vectors and sends designs to a Brother Skitch over Bluetooth. The AI part is the interface: an agent panel runs Claude Code or Codex against the studio's own MCP server, and while the agent works, the objects it is touching get a moving outline, a shimmer and a label saying what it is doing.

That shimmer is a real design pattern. When an agent edits your canvas, show where and what, live. More tools need this.

7. Modern Web Guidance: teaching coding agents the modern web

Paul Irish and the Chrome team shipped Modern Web Guidance 1.0 on November 10, made with Firefox, MDN and Edge. It is an agent skill and CLI with more than 200 hand-written guides, built because coding models keep reaching for old div-and-JavaScript patterns instead of what browsers now do natively, like <dialog> with proper focus handling or field-sizing for textareas that grow as you type. The Show HN thread passed 430 points.

Not a flashy creation, but it is the one I will use. The UI an agent writes is only as good as the patterns it defaults to, and native patterns are usually the accessible ones.

8. Yogurt: private meeting notes that never leave your laptop

Eric from Frigade posted Yogurt on November 11, a free, MIT-licensed desktop app for meeting notes. It records calls without a bot joining, transcribes with whisper.cpp (and Apple's speech framework on supported Macs), and summarizes with an open-weight model through llama.cpp, all on your machine. A local MCP server lets you hand notes to an assistant only when you choose. The Show HN post passed 290 points.

The design call here is the default. Private first, cloud only by permission. As I learned at ViewStats, show value first and ask for trust second.

This week's AI creations at a glance

Creation	Creator	Built with	Why it matters
Saltline	Felix Rieseberg	Claude Fable 5.2	A full 3D game from a two-sentence prompt
Aiden's Revenge	Lex Toumbourou	Claude Code, Fable 5.2, ChatGPT Images, ElevenLabs	A child's brief, shipped as a 12-stage game
Three gadgets	Anthropic, with Diode	Claude, Diode, KiCad	AI moving from screens to circuit boards
Maya site	Pietro Miotti	48 AI agents, public lidar	Fast looking, weak at prior work
Tiny València	@13ug1mb	Real 3D data plus a fine-tuned image model	Real geometry, AI only for style
ZigZag	Kath Korevec	Rust, MCP, Claude Code or Codex	Showing an agent's work live on a canvas
Modern Web Guidance	Google Chrome team	Agent skill, 200+ guides	Better default UI patterns in agent code
Yogurt	Frigade	whisper.cpp, llama.cpp	Private by default AI notes

From my lab

Each week I add my own experiments. This was a week of games and sound, so here are two of mine.

stdo.: a landing page you can play like a keyboard

In my stdo. experiment, every key on the page plays a sound and you can record a loop with a step sequencer. It is built with Figma, Claude and the Figma MCP, and the sound is Web Audio synthesis generated in the browser, not audio files. The experiment page links the playable demo. Aiden's Revenge leaned on ElevenLabs for about a hundred sound effects; stdo. goes the other way and makes the interface itself the instrument. I wrote more about why sound is the layer most teams skip in my post on UI sound design for product designers.

Dice Board: a board game built in an afternoon

My Dice Board experiment is a native iOS board game: a real 3D die that rolls, and a token that hops tile by tile grabbing coins. Tools: Figma, Claude, the Figma MCP, SwiftUI, SceneKit for the 3D dice and Higgsfield for the visuals. The experiment page has a video of it in action. Next to Saltline it looks tiny, and that is the point: the fun is in how the die lands and how the token hops, which is a design decision, not a token budget.

What this week means for designers

The loud story this week was model news, again, with Fable 5.2 and Gemini 4 Hydrogen, plus a lot of noise in tech that had nothing to do with design. The creations above are what you can learn from. They share a shape: the AI handles production at a scale that was absurd a year ago (a shooter, a 12-stage game, a four-layer circuit board), and the human work moves to the edges. Who writes the brief (a nephew). Who checks the facts (Pietro's update). Who decides what the model is not allowed to invent (Tiny València's real buildings). Who designs how the agent shows its work (ZigZag's shimmer).

The hype check. Saltline's eight hours and $550 are Felix's own figures, and he works at the company that makes the model. "One-shot" in Aiden's Revenge still came with playtesting and follow-ups. The gadgets are uncertified prototypes. The Maya find was a rediscovery. None of that makes them less interesting, but it should stop anyone telling a client that one prompt replaces a team.

My take, which I also made in my Fable 5.2 post on UI design taste: the model can produce a good-looking default, but the specific, surprising choice still comes from you. At ViewStats the hardest part of shipping AI tools was never the interface, it was trust, which is the story in my ViewStats creator tools case study. The people who do best in weeks like this are designers who can open the code and change it, which is what a design engineer actually does.

A short checklist if you want to try one thing from this week: write a brief as specific as a kid's (name the bosses), let an agent build it, then spend your time on the three things it cannot judge for you: what it should never make up, what the player feels in the first ten seconds, and how it sounds. If you want that kind of work inside your product, with someone who designs it and ships the code, work with me as your design engineer. Last week's issue, with glasses that read silent speech, is the November 6 roundup of AI creations, and more of my builds are on my AI experiments page.

Frequently asked questions

What were the best AI creations this week?

For November 7 to November 13, 2026: Saltline (a browser shooter Felix Rieseberg says Claude Fable 5.2 built from one prompt), Aiden's Revenge (Lex Toumbourou's 12-stage, four-player beat 'em up for his nephew), three working gadgets with circuit boards designed by Claude, Pietro Miotti's 48-agent lidar search that rediscovered a Maya site, Tiny València (an isometric pixel-art map of a whole city), ZigZag (open-source embroidery software with an agent panel), Chrome's Modern Web Guidance for coding agents and Yogurt (private, on-device meeting notes).

What is the best AI this week for building games?

The model behind the most-shared builds this week was Claude Fable 5.2, released on November 9: both Saltline and Aiden's Revenge used it. Aiden's Revenge also used ChatGPT Images 2.5 for art and ElevenLabs for voice and sound effects. Those are the creators' own reports, not benchmarks, and both still needed a person with a clear idea of what the game should feel like.

What AI tools are people using to build apps and websites right now?

This week the named tools were Claude Code and Codex (inside ZigZag's agent panel), Claude Fable 5.2 for long autonomous builds, Diode and KiCad for hardware, whisper.cpp and llama.cpp for on-device AI in Yogurt, and agent skills like Chrome's Modern Web Guidance to improve the front-end code agents write. In my own experiments I mostly use Figma, Claude and the Figma MCP, plus SwiftUI, Web Audio and Higgsfield.

Can AI really build a whole game from one prompt?

Sometimes, with caveats. Felix Rieseberg says Saltline came from one short prompt and eight hours of autonomous work, at about $550 in API costs. Lex Toumbourou's game started from one large prompt but got extra features after playtesting. A playable result is now realistic; a polished, tested game still needs a designer making calls.