Claude Code 2 and the hidden cost of slow coding assistants: context switching

The ironic part is, that i often find human readable changes (or slightly faster) better for coding, then a LLM that spends x time thinking and then dumps all the information/changes in one go.

It allows me to monitor the progress, pauze/stop the changes, maybe undo a bad change, and instruct it to fix or add context.

Having a LLM that is ultra fast sounds great, until you need to undo a ton of bad changes. Programming even with a LLM, is still about understanding the changes being done to the code base.

If you only program in non-attenuative mode, and act on the total sum of the changes, you need to be darn sure that your start prompt is very detailed to anticipate the LLMs behavior.