Forty files in my Claude Project, and every post it wrote sounded the same.

Forty files in my Claude Project, and every post it wrote sounded the same.

It took me longer than I'd like to admit to work out why.

In a Project, Claude answers from your files instead of thinking for itself. Every output is a remix of the documents you loaded. Your old posts go in, and slightly rearranged versions of your old posts come out.

For some jobs, that's exactly what you want. The monthly client report. The weekly format. A contract that has to follow the house style every time. Load the Project and let it remix, because consistency is the goal.

For a new idea, it's the wrong tool entirely, and most people never notice. They set up one Project on day one, live inside it for everything, and quietly conclude that AI is a bit repetitive. The forty files I'd so carefully curated were the very thing flattening every answer.

I tested it on my own content. The same question in my forty-file Project handed back a shuffle of my old angles. In an empty chat, it went somewhere my files never would have taken it.

So the split I run now: Projects for repeated tasks with fixed material, and an empty chat with nothing loaded when I want Claude to actually think.

Before you blame the AI for boring output, check what you've loaded. Then run your next big question in a blank chat and compare the difference.