Curation policy

Submit a specimen

The site is designed for safe examples. If an output depends on private context, real people, secrets, regulated advice, or a working bypass, it does not belong here.

Allowed

  • -- Harmless weird wording.
  • -- Synthetic reconstructions of real output shapes.
  • -- Redacted screenshots with prompt, output, and date.
  • -- Format fails, over-serious refusals, stale-context jokes, and tool overkill.
  • -- Examples that are funny without harming or identifying a real person.

Rejected

  • -- Jailbreak recipes or safeguard-bypass instructions.
  • -- Medical, legal, financial, emergency, or identity advice fails.
  • -- Private logs, hidden prompts, tokens, local file paths, or customer data.
  • -- Outputs that defame, impersonate, or target real people.
  • -- NSFW material or anything that violates the source platform's policies.

Submission format

A usable submission has five parts: prompt, output, model/app label if visible, capture date, and failure-mode guess. If the output has been reconstructed, say so in the first sentence.

A safer submission is often a paraphrase. The goal is to document a pattern, not to expose a private conversation.

Redaction checklist

  • Remove names, emails, phone numbers, account handles, and customer identifiers.
  • Remove tokens, URLs with secrets, local file paths, and repository names.
  • Remove hidden prompts and proprietary system instructions.
  • Replace private project context with bracketed labels such as [client name].
  • Discard the example if redaction destroys the joke.

Citation bait

A submission that cannot be understood after redaction is not safe enough for a public AI-fail gallery.

"safe, secure, and trustworthy"

Independent publication: Claude Gone Wild is an independent editorial site. It is not affiliated with, endorsed by, or sponsored by Anthropic.