You’d think that with more context it would get better, but it doesn’t.

My hunches are that the to-and-fro of ideas as you discuss options or make corrections leads to a context with competing “intentions”; and that they can’t tell the difference between positive and negative experiences when it comes across each successive token in the context.

But I don’t make LMMs, so this is pure guesswork.

Quite honestly after using Copilot, Claude Code, Cursor, and Codex extensively for months, I very rarely hand edit code.

The rabbit hole problem is pretty rare. Usually it happens when the model flips into "stupid mode", some kind of context poisoning. If you are experienced, you know to purge the context when that happens.

In personal projects I avoid manual editing as a form of deliberate practice. At work, I only edit when it is a very small edit. I can usually explain what I want more concisely and quickly than hand editing code.

I probably would use more hand editing if I had classic refactoring tools in the IDEs similar to intellij/pycharm. Though cli based tools were a pleasant surprise once I actively started using them.

> The best use is editing code with it instead of purely relying on prompting.

What does this look like in practice?

This can go both ways though. I've had instances where after a long somewhat tedious back and forth with Claude Code what I'm looking for finally "clicks" and it starts one-shotting requests left and right. But then after coming up to the end of the context window, I compact or ask it to update CLAUDE.md with what it learned and it loses the nuanced understanding it built up through trial and error and starts screwing up in the exact same ways again with fresh context.

And then I waste 20 minutes bashing my head against the wall trying to write three paragraphs meticulously documenting all the key gotchas and lessons from the "good" context window with the magic combination of words needed to get Claude back in the right head space to one-shot again.

At least if I pull that off I can usually ask Claude to use it as documentation for the project CLAUDE.md with a pretty good success rate. But man it leaves a bad taste in my mouth and makes me question whether I'm actually saving time or well into sunk cost territory...