The whole “Code Output” section is horrifying especially with how I have seen Claude operate in a large monorepo.

This mode of operation results in hacks on top of shaky hacks on top of even flimsier, throw away, absolutely sloppy hacks.

An example - using dict like structs instead of classes. Claude really likes to load all of the data that it can aggressively even if it’s not needed. This further exhibits itself as never wanting to add something directly to a class and instead wanting to add around it.

The best way to approach these (imo) is to pick out some things you think will be helpful. It's a giant vibe fest on this front since there is little in the way of comprehensive evals and immense variation in what people do. Having iterated a bunch on the tone / output formatting, it doesn't seem to impact capabilities (based on my vibe-vals)