Claude Code Builds Sophisticated SimRefinery Prototype From Screenshots

Here is the old GPT-5.1 version in contrast: https://emollick.github.io/simrefine_reconstruct/

@acatovicx Nothing

@emollick Whoa

@emollick What did you use in terms of graphics assets and libraries?

@emollick I'm going to try this for a modern SimAnt or SimEarth.

@emollick How much was the token cost on this?

@emollick What a time to be alive

@emollick Can't wait for fable 6

@emollick this is going to be a major hit in houston!!

@emollick App stores are gonna be flooded with masterpieces like this, once Fable-level game-dev becomes as cheap as Opus or Sonnet.

@emollick the 10 month gap is the real story here. stuff i gave up on last year because the model couldn't hold the whole thing together, i should probably just retry from scratch. reconstructing from screenshots alone is wild

This is exactly what excites me! I built a physics simulator with Fable this week for my students and the learning mode is the bit that matters - it turns a demo into something they actually explore. My concern though: would these hold up when a student tries to break them, or just on the 'happy path'?

@emollick Screenshots and docs as the entire spec is the wild part. Ten months ago this needed hand-holding, now it adds a learning mode unprompted. Software archaeology might be the best agentic coding benchmark yet. What did it still get wrong?

@emollick THIS, is a fair comparaison :) I need more comparaison like that to understand the gap between models

@emollick bro no fucking way this is so lit ill have you know i love sim games and this shit right here is rad

@emollick Great AI training for drone activity both in the Persian Gulf and Russia 🤯

@emollick so far only complaint is holy crap the timer goes fast. you just have to pause it to do anything.

@emollick Love sim games! Is this on github?

@emollick The SimRefinery rebuild shows how much frontier models rely on latent knowledge vs. pure reasoning. I'd test whether it fails on niche domains without prior documentation that's where most production workflows break.

@emollick Reconstructing a lost Maxis simulation from just screenshots and documentation into a fully playable game with a learning mode is wild. The 10 month difference in output quality says everything about how fast this moved.