Hello everybody,
I read posts about the thinking problematic by some model.
I encountered the same problem very recently by applying prompt instructions that very simply extract content from given web site pages, sumarize the content from the original content and put it in a json file dedicated part in a given location.
I think the manual operations by using IA with no agentic magics takes 30 seconds because I have to do some editing manually.
The execution by Cursor and IA using my prepared prompts took minutes however to finally get the same final result 100% succesfully.
Why this difference ?
In the manual case I only ask IA to summarize the original content and I put the result into the file at the right place in the file I created. I will launch the tests by my self (cucumber jest tests that validate the file regarding the json schema).
In the case I ask Cursor to do everything : creating the file by using json shema structure, get the content from the web site, summarize it, read and write the json file to put the result into, launch the test to validate everything, I have opened a complete different context window I think
- more .cursor/rules
- more tools and more tools call
This is what I think it is doing the model think too much.
The problem is that I don’t know the part of thinking I can avoid with a better prompting that instruct clearly the already validated way to achieve tasks in the given context because I don’t have access to the iterative cursor agent execution plan that is briefly displayed in the chat but seem not stored in logs. These informations would allow a bottom up optimization focused on the most critical part on the execution plan (as I will do it in SQL by example for poor performance requests).
I prefer this approach rather than searching the magic prompts that executes better !
Is it a way to get these informations ?
I don’t think so and I can imagine some IP protection reason.
It’s a challenge for Cursor, because the effectiveness of the usage of agentic is clearly in allow us managing correctly the execution plan to limit the context window and the calls.
Regards