How to prevent Cursor from auto-pulling related files? Token usage is crushing my allowance

Support Help
1

Hi everyone,

I’m working on a large project and my token allowance is getting demolished, cache read tokens is in millions+. I’ve read about using `.cursorignore` and manual context with `@File`, but I’m still confused about a few things:

  1. Does `.cursorignore` actually prevent Chat/Composer from accessing files, or does it only affect indexing? I’ve seen conflicting information about this.

  2. What’s the most effective way to work with ONLY the specific files I manually select, without Cursor automatically pulling in related files as context?

  3. For those working on large codebases - what’s your workflow? Do you start a new chat for every single task, or is there a better approach?

I’ve already tried:

  • Starting fresh chats frequently
  • Using `@File` to manually specify context
  • Creating a `.cursorignore` file

But I’m still burning through tokens faster than I’d like. Any tips from people who’ve solved this problem would be hugely appreciated!

Just subscribed to pro+ and hit the limit in one day.

Thanks in advance!

2

Just ask Agent to work only with linked files and request permission if it needs to get additional context.

But maybe you have a problem with context engineering: which files you provide to him in the prompt. He does not have enough data to make decisions, and he begins an extensive search.

3

Understood, maybe this can be the case, not suficiente context to the agent take the necessary action.

It’s scary to see cache reads skyrocket, millions each request.

What’s your workflow? how you set boundaries to prevent this behavior? Are you using the agent to build full features? reviews? refactors?

4

99.9% of the code I write is generated by an LLM. However, I mostly write personal projects, or stuff that’s just sitting in a folder somewhere (hopefully just for now…) or on my GitHub profile.

I use my Agent Compass and I also set up a local CI, like Agent Enforcer, in every project so the AI can self-check before submitting its work. But either way, I rarely manage to oneshot a task.

Maybe I should update Agent Compass with my prompting notes :thinking:

I’m gonna risk it and say you’ve got some janky project architecture. Agent has to slog through way too many different files just to find one specific thing. Are you even following SRP?

I haven’t even finished one commercial project yet, but in all my own repos I start getting twitchy when I see files with four-digit line counts, and I cover the project with 75% test coverage just to guard against regressions.