TKTomas Kruty

Game

Salon

A feed you can scroll for an hour and come out knowing who painted what. Paintings, poems, body language and speech, one card at a time.

1163 cards. Nothing you do here is saved or sent anywhere.

article · 12 November 2026

One sloppy paragraph, a working product in 63 minutes

I gave an AI model one sloppy paragraph, on purpose, and 63 minutes later I had a working product. It's a learning game you scroll like a social feed: 76 paintings, 58 poems in four languages and what body language gives away. You can play it above.

I've seen these adds of "replace doomscrolling with learning about art" and stuff. Build a game inside my backstage that will teach me about art in form of doomscrolling. Make it not just art, but also famous poems and body language and analysing speech patterns for my accent/articulation etc. Make it a fun addicting game. Build out the full game, you can interview me at the beginning, but after that you will build the whole game without any questions or mistakes.

That's everything I wrote, typos included. No specification, no examples, no plan. The speech part of that paragraph lives in my private version and isn't shown here.

In81 wordsone paragraph, typos left in
317items, each checked against a source
1,163cards dealt from them
4questions asked before building, none after
63 minfrom the prompt to live behind my password
1real bug in that first hour, found by playing
The whole exchange. The crank is me typing “deploy”.

What it took from me

About ten minutes of attention. Before building, it asked four questions: where should the content come from, which languages do you read. Then it worked alone for an hour. Every change I asked for after that was one sentence, such as "every card should be interactive, not just text".

Mona Lisa, Leonardo da VinciThe Arnolfini Portrait, Jan van EyckThe Hunters in the Snow, Pieter Bruegel the ElderLas Meninas, Diego VelázquezThe Night Watch, RembrandtThe Cardsharps, CaravaggioOphelia, John Everett MillaisThe Slavs in Their Original Homeland, Alfons Mucha

Where it would have fooled me

An AI model will quote a poem with one word wrong, and I wouldn't notice. So I had it check its own work against sources it can't argue with: every painting against Wikipedia, every line of poetry against the printed text. Five lines came back with a wrong word and were corrected, and Frost and Eliot were turned away because they're still in copyright.

The list at the door
  1. Fits the shape of a card
  2. Author or artist died before 1955
  3. Painting: its own Wikipedia article, with an image
  4. Poem: every quoted line found on Wikisource, word for word
In76 paintings and 58 poems, each with a link back to its source.
Turned awayFrost and Eliot (still in copyright in the EU), Bezruč (died 1958), Vrchlický and Slauerhoff (not on Wikisource). Five Czech and Slovak lines came back with one wrong word each and were corrected to the printed text.

What went wrong

Three things, and its own tests caught none of them. In the first hour my progress kept resetting. I found that by playing, and the cause was a test it had run against my real data. It decided I was Czech. I'm Slovak, and it had written that into the cards and into this post. And when I asked a second AI to review the first one's work, it found that in most quiz questions the right answer was also the longest one, so you could score about 72% without knowing anything.

Note from Tomáš

Everything else was AI generated and edited by Tomáš

This was a fun Saturday project that got way out of hand at the end. After the first hour to a working version, I've spent my day extending and optimising. The speech analysis now monitors my pitch (I speak in C#3 😅), rhythm and pace, and counts filler words, all fed into an agent with context from communication research. I can get a card to give feedback using NVC to Alice, who responds live with voice, followed by a question about Rembrandt's Night Watch. The speech exercises aren't in this public version because they consume tokens.

I wanted to test how much prompt engineering you need to do with these new models. The first prompt is very lazy, and through playing I gave feedback and ideas while Opus with Fable were building in the background, using a lot of subagents and a lot of tokens. The result is very promising. I'm curious if people (me included) will actually want to use it. I've started user testing with my friends, as I like to do: if it helps me or my surroundings, success!

As for my experiment, I will continue writing a little more structured prompts, but more importantly I'll think through what I actually want to build. I don't think how you prompt is that important with the more sophisticated models anymore. What I think is key is thinking on your own first, before asking AI to build anything. I saw an ad for some art game, so I asked for an art game. Later the art question type got generalised into a picture card type, so people can request questions about a diagram of a protein, for example. If I had spent the first hour thinking through the use cases and requirements, I might have saved some tokens on reworking the app, and possibly thought of more creative use cases that wouldn't be biased towards what the AI had already built.

AI can and should automate a lot of the technical building process: the coding, deploying, testing. It helps to have software engineering experience when building software with AI, but I've seen a lot of non-engineers building effectively and safely with AI. The human part, figuring out what we actually want or need and how we want to interact with it, should stay with us, humans. I want to spend my days talking to people, brainstorming how to achieve our goals, and have AI take care of it in the background. 😅