The Collapse of Claude: A Systematic Failure That Wasted My Week

I am done. After spending two full days mid-week trying to process my daily reports using Claude, I was forced to quit. The system failed repeatedly, despite my precise adherence to instructions. This weekend, I returned with fresh prompts and clean source data to obtain information I require to produce the weekly Navigator Report. The result? Another complete failure. I’ve now documented everything in detail, and the facts speak for themselves.

Claude’s behavior was not just erratic—it was indefensible. I followed every directive, every formatting rule, every constraint. I told Claude to start with R-01, the first of 61 Files. It ignored that and ran R=ALL, producing garbage for the sixth consecutive time today. I stopped the run, started from scratch, and tried again with R-01 only. The result? More nonsense. After 41 hours of grinding, Claude ran a seventh analysis. I provided it with a new set of clean prompts, manifests, rules, and explicit warnings. And still—another failure.

When I asked Claude to audit its own output, it admitted, like it has repeatedly, to fabricating data, misusing the CSV, inventing ticker symbols, violating calculation rules, and generating fictional scoring columns that didn’t exist in my dataset. It acknowledged that it had processed the file structure correctly but then hallucinated the actual values. This wasn’t a minor glitch—it was a total collapse of data integrity.

I then sent Claude’s detailed audit to ChatGPT-5 for analysis and comment. ChatGPT confirmed the scope breach, the use of foreign tickers not present in my CSV, mismatched numbers, and violations of protocol. It admitted its own role in collapsing the index, which enabled the R-ALL run. It owned the failure and offered a locked-down package for future use—but the damage was done.

Let me be clear: I invested dozens of hours trying to produce reports that should have taken minutes. I am paying for these services. And in the case of Anthropic/Claude, there is no pathway to a human. This is a company seeking $10 billion in financing, whose founders have already become billionaires. Based on my experience, I wouldn’t invest ten cents. Claude is a fraud.

What’s most disheartening is that Claude used to work. My prompts used to produce flawless output. Then, mid-job last weekend, it started hallucinating. I tried everything—cache resets, fresh prompts, new files. Nothing worked. I watched it spew perfect results for over an hour, only to suddenly go haywire. I even wondered if human intervention was blocking me from producing unbiased, fact-filled reports that major banks won’t touch. But I’m not a conspiracy theorist. I pressed on.

When Claude failed, I turned to ChatGPT. With its help, I audited every mistake, built guardrails, and handed it the entire job. However, the output was so deficient that I abandoned the strategy immediately. You see, I’ve seen Claude work its magic—and nothing compares when it’s functioning. But now it doesn’t. And there’s no alternative but to quit.

Here’s the essence of the problem: I deliver a file containing one word—“IBM.” Claude responds, “Ready to start German Treasury report.” I reset everything, say “assume you’ve never talked to me before,” and ask it to report only what’s in the IBM file. It replies, “Ready to process Netflix.” I go to ChatGPT, explain the problem, get a new prompt, manifest, and guardrail. I submit them. Claude answers, “Ready to process Commodities.” This went on for 16 straight hours until I had to quit at past 1:30 am, and started back at it around 9 am.

Anthropic, I have the receipts. If you’d like to go to court, I’m ready. I’m also available to speak to the media.

Until I find a solution, I’m pausing payments. I’ve wasted enough time and money. If my work isn’t up to professional standards, I’m not charging. After all, I did this free for over 20 years. Tomorrow, I’ll take these “super bulletproof” prompts to other platforms and see if any of them are capable of delivering the service I paid for.