Fix the folders, not the robot

The other day I was driving home from work, turning over a question I keep circling: what's the best way to teach ICM, or to learn it? Almost immediately it clicked. Make it a game.

So I pulled out my phone, opened the Claude app, and we went back and forth until the concept held together. I wrote it up as a PRD, and the minute I got home I handed it to Claude Code. What came out is a little game to teach ICM, the methodology we owe to Jake van Clief: an agent only knows what's on its path. Get the structure right and it does great work on its own. Get it wrong and it fails, cheerfully.

Meet Bip. Bip is a robot intern at a painting shop, and Bip is not stupid, just literal. It reads exactly what's in front of it and does exactly what the signs say. Your job isn't to fix Bip. It's to fix the folders Bip has to walk. Move the right file onto its path. Drop a signpost in the room where the work actually happens. Send a helper into the messy folder so Bip doesn't drown in it. Wall off the stages so yesterday's notes don't leak into today's invoice.

Thirteen levels, about fifteen minutes. Each one breaks in a different way, and every break is a real ICM lesson you'll feel before anyone explains it. Watch the battery drain and the little brain jar fog up when the workspace is a mess. That's the cost of bad context, made visible.

The best bit is the end-of-level screen: your first messy run next to your clean one. Same job, same robot. The only thing that changed was the structure. That's the whole idea of ICM in one picture.

Credit where it's due. The method is Jake van Clief's. This is just a way to play with it.

Have a go, then tell me how I can improve it!

https://claude.ai/code/artifact/52eee8c6-a5c2-401c-a2a0-8630848288ef

8 comments

Fix the folders, not the robot