Observation: edit block takes a lot of token to repeat code that needs to be replaced

[xCatG]

I wouldn't call this a bug but more like asking for opinions; I observed that when editing block because LLM need to spit out the exact block that needs to be replaced, it caused a lot of wasted tokens while doing so.

However LLMs are also very bad at generating proper diff format due to they can't count correctly, so I don't have any alternatives to suggest. Do you know any methods for code replacement that is more accurate yet doesn't need to repeat tokens that need to be deleted?

[wonderwhy-er]

You are correct. For that reason its prompted to use full file rewrites if it intends to rewrite a lsrge block. Line based and diff based and even regex based replace are way less reliable for reasons you mentioned.

Models are good at writing text in that sense.

[fezzzza]

The shorter the chat, the more accurate Claude is at being able to do diff block edits. I restart new chats as often as possible as soon as a minor task is complete. It saves tokens and stops me running out of tokens before the five hour window expires.