I sent Claude Opus 5 '–-' and it wrote me 5k tokens about a cartographer

Three dashes, one AI meltdown, and the comments are already at war

TLDR: A writer found that sending Claude Opus 5 only three dashes could make it produce huge, bizarre stories instead of asking what you meant. Commenters instantly turned it into a fight over whether the bug was amazing, overhyped, or just proof that too many users are new to chatbots.

A simple string of dashes somehow turned into full-blown story time, and the internet immediately did what it does best: split into camps and start yelling. In the original post, the writer says that sending Claude Opus 5 nothing but markdown-style dividers like --- sometimes made it spill out thousands of words about random topics, including a cartographer and a woman named Marisol. Instead of asking for clarification, the chatbot allegedly acted like it was mid-sentence in somebody else’s conversation. Naturally, this sent curious users running to test it for themselves.

That’s where the community drama kicks in. One side was fascinated by the absurdity: people were basically treating the chatbot like a haunted vending machine — insert punctuation, receive surprise novel. Others were far less charmed. One commenter, pavlov, threw a direct jab at the post itself, calling it “obviously AI-written” and saying it was hard to follow without missing background. Ouch. Meanwhile, Oarch came in with the classic buzzkill reality check: they tried it themselves and got a boring response asking what they meant. Translation: not everyone got the magic dash jackpot.

Then came the veteran-user flex. Tiberium basically rolled their eyes at the whole spectacle and argued that anyone shocked by this should spend more time using simpler text-generation tools. So the mood was equal parts “LOL the robot is writing fanfic from punctuation” and “please learn how these systems work before acting surprised.” In other words: a perfect internet mess.

Key Points

  • The article says Claude Opus 5 can sometimes respond to `---` with long unsolicited outputs such as stories, essays, or other documents.
  • The behavior was first highlighted publicly on X through posts showing related failures involving dash-based separators.
  • The author reports running 649 API calls on AWS Bedrock and 720 on the Anthropic API to characterize the issue.
  • One cited example is a 10,601-character story titled *The Weight of Wings* generated after the prompt `---`.
  • A delimiter sweep described in the article suggests the trigger is linked to markdown-like grammar rather than arbitrary empty input.

Hottest takes

"obviously AI-written" — pavlov
"Looks like your message came through empty" — Oarch
"People should just try base models" — Tiberium
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.