The models or cursor itself seem to be getting more and more eager to do more work, but there is a cost to that. For example, when I request a dependency analysis on a table, I am happy with a brief summary of the classes and methods that might be affected by a change, but I get pages of response, plus a visual dependency diagram. I don’t really need that. Another example is when the model goes off on like “oh hey, I need to check this too, let me add some common methods here you didn’t ask for, and then and let me run linter on the whole code base now even though I only changed one file”. I appreciate the effort, but a lot of it is busy work and comes at a cost which we are all trying to control. My request is some kind of configurability where we can choose the general level of “Due dilligence” happening with a targeted request.
hi @kev_rm this is best done with “Show me only a brief summary” as we may not know what kind of level of detail each user or task requires.
@condor sure but my suggestion was more about the tool is clearly biased towards costing the user more. verbose should be opt in, not the other way around. at a minimum make some kind of middle ground or setting.
Thank you for the suggestion. AI model behavior is set by AI provider during training or otherwise controlled by a user prompt, meaning you can also tell AI what you want done in which way.
Once it becomes a setting we can use on majority of APIs we will consider it.
@condor just to be clear: The models are.. frequently, like 80% of the time, doing things far beyond what is was asked to do. You get paid by how much work the model does.
I run out of tokens. Can I get back my tokens for stuff I didn’t ask for?
It is reasonable to want to control costs, but he is saying that is how the models are built from the provider (OpenAI, Anthropic, etc). You just have to learn how to manage the tool (model). You don’t tell a mechanic: “just fix my car” when what you really want is “fix my car with the least repairs as possible to get it running”. And yes the providers may want more verbose responses because it earns them more, but that’s the provider, not Cursor, so just add a Cursor rule (which is like text concatenated to every request) that says “be brief always… unless I specifically say to give more detail” Problem solved. The providers may also be biased towards more detailed responses because it flexes what the model can do and that is good for their product, especially for first impressions which is why its the default model behavior.
If Cursor started changing the default behavior of these models from the provider then people would complain that they are getting an altered experience on Cursor vs going straight to the provider.
Some people want the detail, some (like yourself) don’t. Luckily you have the freedom to customize the responses the way you want. If Cursor implemented something like you describe, I wouldn’t use it. I want more control than whatever Cursor’s “verbosity” setting is doing.
I have Cursor rules about not creating example or readme.md files because I was annoyed by that extra work which I usually just deleted, but I could see for some people it would be a pleasant addition. Its very specific to the user.