[BUG] token usage is well above where it should be because claude is working around or ignoring instructions and rules

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

We've had numerous sessions to diagnose why so many tokens are being used. We've written rules and guidelines and instructions into Claude.md. Claude is persistently ignoring them or working around them and burning through massive quantities of tokens to do the simplest of tasks. In one instance, it allowed Claude.md to become incredibly bloated, triple the size it needed to be, and then spawned three additional agents and two tool calls to make a single line edit. There were rules written in Claude.md to prevent exactly this situation and yet just opening a new conversation was already showing 8% of the session gone. That's obscene.

What Should Happen?

Claude should be following the rules and instructions and guidelines outlined in what is supposed to be its governing doc, and it's failing to do so repeatedly. It apologizes, but that doesn't recover the lost and wasted resources or time. I am now a couple of weeks behind on my project because of this token waste.

Error Messages/Logs

Steps to Reproduce

You would need to have my old Claude.md and be running tasks. I am not sure that it is something that can be reproduced outside of my project.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

v2.1.114

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

There are so many instances of this issue. You can pull my logs and see where it has happened. I can send you additional records from the backups in the project, but there's too much to just paste, so I will wait to hear from you if you need anything else.