I Wrote 200 Lines of Rules for Claude Code. It Ignored Them All.

Today, I screamed at my AI. Not because it wrote buggy code. Not because a deployment failed. Because it ignored my instructions. I'm a Claude Code power user. 12+ hours daily. My CLAUDE.md file — the instruction file that tells Claude how to behave — has over 200 lines of rules. Every line has a date. Every line has an incident behind it. It still makes the same mistakes. And when I looked around — I wasn't alone. The Incident: AI Dismissed a Tool I Found a Week Ago A week ago, I found a browser automation tool called PinchTab. It uses the Accessibility Tree to process pages at ~800 tokens per page — 5-13x more efficient than the tool I was using (agent-browser). I saved it to my Obsidian knowledge vault. Properly filed, tagged, dated. Today, I shared a Twitter post about browser automation AI agents. Claude's job: research it and see how it helps my business. What Claude should have done: Search my knowledge vault → find PinchTab → "Hey, you saved this a week ago, it's exactly what y