Hey. I couldn’t figure out a way to do the following in the current state: with many agent requests, I care mostly about the code generated and want to review that. I don’t care so much about the explainations (only in maybe every 5th request). Would it be possible to somehow narrow down the model response to the very essentials and keep additional infos out of my view? I know its possible to set a cursor rule and ask the model to stick to the essentials unless asked, but that would create additional requests that I don’t think are ‘worth it’. Maybe you are already working on something that caches or hides the explainations and gives the use control over the detailness of explainations? For reference, I’m using Claude in different versions. –Arwed
I usually split tasks: First I tell the model to create an outline of proposed changes, with a brief rationale. Then, I either accept that or engage in conversation about the proposed changes. Only once I am satisfied that the approach is good, I tell it to code.
But you are right: it is way too talkative. I’d like to see a bit of explanation/reasoning interspersed with the actual code, and the possibility to interact with each step to get a more in-depth exposure to the reasoning.
All this bla-bla by the models is not just cluttering my feeble human mind, it also spams the context window. To the very least, devs, can you cut out on all the “polite” BS? “You are right…I apologize…your approach is better because…” Well, yes, I know my approach is better. Cant those models not behave more robot-like? A simple agree or disagree would be fine. I can always ask the model to expand on that.