I Caught Claude Code Breaking Every Rule (9 Times in 1 Hour)

From AngelSix

2026-05-28 18:05 · Open original ↗

I spent the entire day fighting Claude instead of shipping code. In less than one hour, it broke my hard-coded rules 9 times — rules that should have been impossible to miss. This is what AI agents actually do in production, and it's genuinely dangerous. In this video I walk through every failure, show you real scraped logs, live demos where Claude ignores its own primer, and — most importantly — the guardrail system I'm building to fix it: a Global MD rule file, non-AI verification scripts, context recall checks, and local model hosting. This isn't theory. This is what happens when you trust AI blindly. ▬▬▬▬▬▬▬▬▬▬ TIMESTAMPS ▬▬▬▬▬▬▬▬▬▬ 0:00 - When AI Lies Instead of Codes 0:45 - My Custom AI Tooling (Built With AI, Barely Trusted) 1:34 - Why We Need Hard Guardrails for AI Agents 2:11 - Rusty Guard Chat: Using AI to Review Its Own Failures 2:34 - Example 1: Claude Blamed Me for Its Own Mistake 4:17 - The Full Log: 9 Rule Violations in 1 Hour 5:05 - Testing New Guardrails — Immediate Failure 5:42 - The Global File That AI Keeps Ignoring 6:23 - The "Rusty Rules Followed" Verification Check 6:44 - Live Demo: Claude Fails to Read Rules in Real Time 8:44 - Mid-Session Context Validation 9:46 - Why Only Non-AI Scripts Can Verify AI Actions 10:38 - Rusty Recall: Testing AI Memory Mid-Session 11:46 - Scraped Log Review: 9 Failures Out of 21 Sessions 12:32 - The Scariest Failure: AI Killed Every Active Process 13:11 - The "Get Sync" Skill — Completely Ignored 14:02 - Guardrails on Top of Guardrails (Still Not Enough) 15:01 - Solution 1: Run AI Locally with OMLX 15:29 - Why Cloud AI Is More Danger Than It's Worth 16:00 - Solution 2: The Global MD Rule System 17:37 - Memory, Self-Reflection & Documentation Sync Rules 19:01 - AI Mecha Tooling Tour 20:30 - The Real Dangers of Blindly Trusting AI 21:12 - My Proposed Weekly AI Dev Series 21:57 - What AI Actually Improves (And What It Doesn't) ▬▬▬▬▬▬▬▬▬▬ LINKS ▬▬▬▬▬▬▬▬▬▬ 🌐 Website: https://www.angelsix.com ▬▬▬▬▬▬▬▬▬▬ NEXT VIDEO ▬▬▬▬▬▬▬▬▬▬ Next up: Back to Avalonia UI — no AI, just real code. ▬▬▬▬▬▬▬▬▬▬ TAGS ▬▬▬▬▬▬▬▬▬▬ #AI #ClaudeAI #SeniorDeveloper

I spent the entire day fighting Claude instead of shipping code. In less than one hour, it broke my hard-coded rules 9 times — rules that should have been impossible to miss. This is what AI agents actually do in production, and it's genuinely dangerous. In this video I walk through every failure, show you real scraped logs, live demos where Claude ignores its own primer, and — most importantly — the guardrail system I'm building to fix it: a Global MD rule file, non-AI verification scripts, context recall checks, and local model hosting. This isn't theory. This is what happens when you trust AI blindly. ▬▬▬▬▬▬▬▬▬▬ TIMESTAMPS ▬▬▬▬▬▬▬▬▬▬ 0:00 - When AI Lies Instead of Codes 0:45 - My Custom AI Tooling (Built With AI, Barely Trusted) 1:34 - Why We Need Hard Guardrails for AI Agents 2:11 - Rusty Guard Chat: Using AI to Review Its Own Failures 2:34 - Example 1: Claude Blamed Me for Its Own Mistake 4:17 - The Full Log: 9 Rule Violations in 1 Hour 5:05 - Testing New Guardrails — Immediate Failure 5:42 - The Global File That AI Keeps Ignoring 6:23 - The "Rusty Rules Followed" Verification Check 6:44 - Live Demo: Claude Fails to Read Rules in Real Time 8:44 - Mid-Session Context Validation 9:46 - Why Only Non-AI Scripts Can Verify AI Actions 10:38 - Rusty Recall: Testing AI Memory Mid-Session 11:46 - Scraped Log Review: 9 Failures Out of 21 Sessions 12:32 - The Scariest Failure: AI Killed Every Active Process 13:11 - The "Get Sync" Skill — Completely Ignored 14:02 - Guardrails on Top of Guardrails (Still Not Enough) 15:01 - Solution 1: Run AI Locally with OMLX 15:29 - Why Cloud AI Is More Danger Than It's Worth 16:00 - Solution 2: The Global MD Rule System 17:37 - Memory, Self-Reflection & Documentation Sync Rules 19:01 - AI Mecha Tooling Tour 20:30 - The Real Dangers of Blindly Trusting AI 21:12 - My Proposed Weekly AI Dev Series 21:57 - What AI Actually Improves (And What It Doesn't) ▬▬▬▬▬▬▬▬▬▬ LINKS ▬▬▬▬▬▬▬▬▬▬ 🌐 Website: https://www.angelsix.com ▬▬▬▬▬▬▬▬▬▬ NEXT VIDEO ▬▬▬▬▬▬▬▬▬▬ Next up: Back to Avalonia UI — no AI, just real code. ▬▬▬▬▬▬▬▬▬▬ TAGS ▬▬▬▬▬▬▬▬▬▬ #AI #ClaudeAI #SeniorDeveloper