The Model Is Its Own Best Collaborator

There's a pattern that shows up once you've used these things long enough — the model stumbles hardest at the very beginning, when it's parachuting into a codebase it didn't build, trying to infer conventions from whatever files you've pasted in. It hallucinates file paths. It invents function signatures that almost match the real ones. It has a confident, slightly wrong mental model of the whole thing, and it will write fifteen confident, slightly wrong things before you figure out what the actual problem is.

But then you notice something. If you start a session clean — let the model scaffold the initial code — and then continue in the same context, it becomes strangely coherent. It extends things correctly. It doesn't contradict itself. It seems to know where the bodies are buried because it's the one who buried them.

Which makes complete sense if you stop anthropomorphizing for a second. The model isn't reasoning about your codebase. It's doing very sophisticated pattern completion. And the strongest prior it has on what comes next is what it already wrote. Its own output is, by construction, the most in-distribution thing it could be asked to continue.

So the haywire-out-the-gate problem is real, and it's not a bug they forgot to fix — it's the thing revealing itself. You handed it a foreign language and asked it to translate. It handed itself a sentence it already started.

The practical upshot is obvious once you see it: let the model write the first draft of anything it's going to touch. Don't bring it into an existing system to do surgery. Bring it in at the foundation, or accept that the first twenty minutes are just orientation tax.

The uncomfortable upshot is that the model you're collaborating with is, in some sense, collaborating best with itself.