I use Cursor and Codex at the same time, everytime Cursor cannot give me the answer in time I will switch to Codex.
I found the Cursor to stick to the fact, if I tell it the program gives me some wrong data, it will try to access the data first, not checking the code first, and if the MCP tool cannot provide a tool to fetch the data it still try to fix it or try some other way to get the data out, but maybe the code bug was so obvious just there
Another example, I ask the Cursor whether the lib can provide an API to allow me to do something, it will try to check the TS type declaration or do some probe against the lib first, but check the doc or check the lib discussion or community Q & A much faster
Thanks for taking the time to write this up. What you are describing comes through clearly, and I understand the frustration. When you can already tell the problem is in the code, watching the agent spend its first few steps retrieving data or inspecting a package feels like wasted time, and switching to another tool becomes the faster path.
The model prefers to verify things directly before reasoning from the code or the documentation. That instinct helps when the data really is the source of the problem, but it works against you when the answer is already sitting in the source. It does respond well to being told where to look first.
When the output is wrong, point it at the code explicitly: “The output is incorrect. Please read the code path that produces this value before retrieving any data, and tell me what you find.”
When you are asking what a library supports, name the sources you trust: “Please check the official documentation and community discussions first, and only inspect type declarations or test the package if those do not answer the question.”
If you would rather not repeat this in every conversation, both instructions can live in an AGENTS.md file in your project root or as a rule in .cursor/rules, where they will apply automatically (docs).
I wonder how much of this is people not using skills or rules. Those are critical to my process and now produce (mostly) predictable and repeatable results. And grilling sessions: make the model ask you when it doesn’t understand and/or would have had to guess.