You are running into the exact wall that forced me to completely rebuild
The issue you are hitting with Claude is that flat file architecture (like spreadsheets) just isn’t designed to act as a relational database. When you feed a flat file into an LLM, you will inevitably hit token/data limits because the AI is trying to read and calculate thousands of rows of text at once.
This is exactly why I moved my data into a true SQL database engine. It doesn’t use an LLM chat interface, but it natively handles all the specific views you just listed (cumulative weekly spending, fixed vs. flexible splits, category isolation) in milliseconds without ever hitting a data limit.
The learning curve to get those insights is basically zero—you just click a visual filter instead of writing an AI prompt.
I know I mentioned my project previously, but since you are clearly pushing the boundaries of what flat files can do, I’d highly encourage you to jump into the sample data sandbox I linked in my Show & Tell post. I am actively looking for power users like you to test the limits of a true database engine.