Tried to convince a business owner that folder structure is the whole game. New video.

New long video is up, and it's basically me trying to convince someone who's never heard of any of this that giving your AI a folder structure makes sense.

The way in is a client story. A business owner running about thirty automations, almost all quietly broken the same way. One had been emailing a report to a dead inbox every day, for months. No error, no warning. The AI kept one copy of a fact in its memory, he kept another in his files, the two couldn't see each other, and they drifted until his setup was confidently doing last year's work.

The fix is the rule this community already lives by: every fact lives in exactly one place, and everything else points at it. If it only points, there's nothing to drift.

The part for this room is what I found when I checked how each tool actually assembles context from the folder. Both walk the same chain: a global file read first on every single run, and invisible from inside your project, then the project root, then the folder you're working in, closest file wins. But they fail in opposite directions. ChatGPT caps the whole chain at 32K and past that it silently drops the deepest files first, the exact ones that were supposed to win. Claude reads everything and just listens worse as it bloats. Different failure, same cure: every instruction file stays a thin router.

https://youtu.be/df6JbGqquNU

Curious how you all pitch this to people who haven't felt the pain yet. The story did more convincing than any diagram ever has for me.

16 comments

Tried to convince a business owner that folder structure is the whole game. New video.