Best practices: Cursor AI Settings for Automated Article Generation Using Web Search

Hello everyone!

I want Cursor to perform tasks with the best quality and in an automatic mode. But how can this be achieved?

The main steps that allow you to create the best file include the following steps:

  1. It is necessary to understand the task
  2. It is necessary to read the structure of files and folders
  3. It is necessary to include a point on searching for the necessary information for the edited/created file
  4. It is necessary to include a point for reflection
  5. It is necessary to create a task for changing the file

and so on - for all files.

As a result, for Cursor AI, it is necessary to create a rule that will indicate that the YOLO mode should be enabled with the execution of items 3-5 for each file.

This will be relevant for ordinary text files.

For program files, it is necessary to additionally include points of static analysis, compilation, and creation of unit and other tests.

Further, instructions are written with an option for such settings.
——————————–

Place the settings.json file in the folder:
C:\Users\<USER NAME>\AppData\Roaming\Cursor\User

All other files must be placed in the project folder.

For the AI rules to work correctly in every Cursor AI chat, you must select:

  1. The Web tool
  2. Rules → web_search_rules_eng.mdc
  3. Rules → yolo_mode_eng.mdc

Rules:

  1. Sequential numbering must not be used — this appears to be a feature limitation. Use unordered lists only.
  2. Provide explicit instructions about which ToDo list items we want to see.
  3. Specify all file operation nuances explicitly.
  4. Clearly indicate the position where new commands should be inserted in the ToDo list.
  5. Model - gemini-2.5-flash

Download:

P.S.
@condor - I am sure that such interesting user rules, which significantly improve working on tasks, would be useful to add to Cursor AI.

Hi @nikitayev and thank you for your detailed post.

Could you please review my points below and let me know if I missed something or if you have issues with this functionality.

Agent does behave like that from my experience.

  • It understands the task (based on info provided)
  • Reads the required code from files
  • if an internal search is required to find all occurrences, it does that.
  • Then it plans the change and performs them.

Auto-run can be setup on user level. Agent has already instructions for the tools available.

For static analysis a .cursorrule file / memory / user rule with core commands suffices.

As agent also does search without @web the main requirement is to have Web Search enabled in Settings.

Overall the connecting parts required are

  • Tell Agent in request what to do.
  • Task details, here I suggest placing them in an MD file. then mention the file in request (dont attach)
  • Mention for agent to plan the task or to plan todos.
  • Do not make too detailed rules for that as behavior works already within Cursor and the rules would use up more tokens.
  • Select the model you like.

You can do a lot of the parts with builtin features. Would appreciate your feedback in case I missed something.

@condor You wrote everything correctly, but practice shows that this doesn’t work as needed, even with the most powerful models like Claude Sonnet 4 and Gemini 2.5 PRO. And if we take a simple model like Gemini 2.5 Flash, it works like a small child and without clear instructions, it’s unrealistic to expect good work from it. I wrote my instructions specifically for Gemini 2.5 Flash and tested them on it. With more powerful models, everything should work even better. Without the todo_write tool, the quality of work also decreases significantly.

Also, the lack of full context size without using MAX mode greatly hinders quality work. In this case, my recommendations also help a lot.

1 Like