claude code logo

I was banned from Claude for scaffolding a CLAUDE.md file – Hugo Daniel

by 

API Error: 400 {
  "error": {
    "type": "invalid_request_error",
    "message":"This organization has been disabled."
  },
  ...
}

A minute I’m €220/month “Max 20x” AI “power user” (Is that even a thing?) Next, I am a disabled non-person “organization.”

Like many of my peers I was using the Cloud Code CLI regularly and trying to understand how far I could go with it on my personal projects. With views and approaches to code, I can now test and verify at a much faster pace. drive it in tmux And let it work while I did something else.

Until I got that reaction in one of these sessions.

"The ban hammer was strong!"

TLDR:

If you’re automating signals that look like system instructions (ie creating scaffolding reference files, or using a cloud to find another cloud’s errors and iterating over its CLAUDE.md, or etc…), then you’re walking on a mine.

Organizations of late capitalism, unite!

My account has been banned! No warning and no response, just that message saying my request was invalid because I’m a disabled organization.

I wasn’t doing anything groundbreaking, in fact, I was asking it to tune a tool I use to create project scaffolding.

Yes, you read that right: Project Scaffolding! Possibly one of the most boring things you can think of doing!

bitch is bitch

So I asked Cloud to update their scaffolding tool so that it could include a CLAUDE.md file with instructions baked in for a particular homegrown framework (coffee
boredom coffee).

I was playing around as a “human-in-the-loop” middleware for these LLM tools. Like seeing one instance of Cloud try to “take over” another instance of itself, and the platform’s security guards mistook it for a riot.

To help understand this, there are three main characters in this story:

  • Cloud A (an example of cloud in the tmux pane)
  • Cloud B (another example in another tmux pane)
  • an inefficient organization (i)

The loop was as follows:

  1. Disabled organization (my) asked Cloud A to update the scaffolding tool with cool CLAUDE.md
  2. I went ahead and started a new project with the tool, opened a cloud there (Cloud B) and asked it to perform a complex task
  3. Whenever Claude B made a mistake, I would go to Claude A, paste the error, and say something like “Hey, Claude B made this error”.
  4. Goto I

Running two instances of CLOUDE, one updated the other’s CLAUDE.md because it had mistakes.

This cycle repeated until I was told I was an inefficient organization.

Banned!

to stop! you are wrong

I just wanted a standard reference file for new projects.

At one point Claude A became somewhat angry with Claude B and started yelling! Writing in en-US instead of en-GB, that is: in all capital letters.

I went to check the file, and it was littered with instructions telling him to do things instead of what Claude B would try to do.

My guess is that this probably led to a decline in the “prompt injection” estimate for the non-disabled organization.

I would have loved to see that AI’s face when it saw its own “system prompt” language echoed back into it.

Or I don’t know. This is all just a guess of mine.

"That hole is red. That hole is me."

“It’s not just bad support; it’s automatic exclusion.”

So I rushed to read their documents. What was going on here?

Made an appeal, which was a link to a Google Docs form, with a textbox where I tried to convince some cloud C at a multi-trillion-quadrillion dollar non-disabled organization that I was not only a human being, but also a well-intentioned person.

I didn’t get any reply. Not even an automatic response. 0 com.

So I wrote in support of them, this time with the help of an LLM from another non-disabled organization.

I didn’t get any reply. Not even an automatic response.

And it’s surprising that people complain about civil servants, eh, wait until you have to deal with one of these expensive machines!

After a few days I received an e-mail:

Credit note from Anthropic, PBC for invoice #…

Yes, the only e-mail I received was a credit note refunding my money.

It’s as if they’re saying, “We don’t want to talk to you anymore, there’s some hidden money here”. But hey guys, it’s not a conversation if it’s only one-sided, and here I am talking to a wall.

I don’t even get the “it’s not you, it’s us.” I just got a credit note.

Ahaha I’m so happy!

I’m glad it happened with this particular non-disabled-organization. Because if by chance this happened to another non-disabled-organization that also offers such tools… I would have been locked out of e-mail, photos, documents, and the phone OS.

AI moderation is currently a “black box” that prioritizes security over accuracy to the extreme.

If you’re automating prompts that look like system instructions (ie scaffolding reference files), you’re walking on a minefield.

conclusion

I got my €220 back (ouch, that’s a lot of money for this kind of service, thanks capitalism). I have recompiled the entire scaffolding project, and reverted all the code that was committed there by Cloud.

Soon I will re-release Boredom with a new angle and approach without Claude’s help. I’m trying to convert it to a JS framework for LLM (formerly LLM, or just LLM, now it has no API). To create and iterate over the single .html files that these tools are now exposing the world to.

If you want to take a look at the CLAUDE.md with which Cloud A was running Cloud B, I’ve committed it and it’s available here.

To wrap it up again: This entire post is just my hypothesis. At the time I got banned, the cloud was doing nothing except replicating this file. And I haven’t heard from them about it anymore (or ever).

“You have to understand that these organizations have a lot of users…”



<a href

Leave a Comment