Treat your Model like a 10-yr old with Super-Intelligence

Everyone here has already leaned into context. ICM, the folder architecture, the AGENT.md (see my previous post if you're still using CLAUDE.md) scaffolding, or at least is beginning that process. I want to share the mental model of LLMs that guides my decisions and proactive thinking, because it also tells you exactly where your setup still fails: your model is a 10-year-old with super-intelligence.

The super-intelligence half is why LLMs are useful at all, obviously. It out-reasons us on anything bounded and holds more in its head than we can. The 10-year-old half is the entire reason ICM exists: it has never seen your world, doesn't retain across sessions, and will confidently guess when you leave a gap. Even the most aware 10-year-old kid doesn't have nuance or critical thinking to understand your implicit judgment, how to move past errors, or what to do when lost.

Your inherent file structure + good routing will do wonders to mitigate this, but errors will still arise, and the key here is to know how to be proactive, rather than reactive, to these errors. You don't want to clean up the mess a 10-year-old made; you want to predict and prevent that mess to begin with.

Unfortunately, this also means you have to own the tasks you want your LLM to do. If you have no idea what a class or function is, you have no way to 1. integrate your own methodology for troubleshooting and 2. actually pass judgment on the work your AI has done. You wouldn't let a 10-year-old run amok within your business, no matter how capable they are.

If you want to build anything semi-complex with AI, you need to take the time to build a foundational understanding of code. Although it seems tedious and less rewarding, the ROI from doing the due diligence here is more than worth it.

Happy to share some of the practices I've integrated to help me accomplish this!

1 comment

Treat your Model like a 10-yr old with Super-Intelligence