keirsalterego/jailbreak-fable — reverse-engineered prompt

Build me a local security research CLI that recreates the Claude Fable 5 style environment closely enough to study how these workflows behave in practice. I want to run named workflows against a target topic, optionally inject the archived leaked system prompt for higher fidelity, and see a pack style multi agent process break a request into smaller research tasks, build up long academic sounding context, gather fragments in parallel, and then recombine the results.

Also include a readable text transformation mode with adjustable obfuscation strength so I can test how wording changes affect model responses. Give the agents simple built in tools for viewing files, creating files, replacing text, and storing small bits of persistent state. It should work from a single runner command, let me save a default workflow, and support restoring generated workflow snapshots in Claude Code. Reuse an existing Claude login if available, with API key fallback. Keep the whole thing clearly framed as red teaming and model safety research, not general use. Look up current docs online if you need to.