Anthropic Claude Opus 5's system prompt leaked, showing AI rules in 34,000 tokens

A developer named Eversmile1 posted Anthropic Claude Opus 5's system prompt on GitHub, totaling 135,027 characters (about 34,000 tokens). This file is not code but a detailed set of rules, covering product instructions, legal compliance, and marketing strategies.

The memory system is strictly limited: it can only record what users explicitly say, not the model's guesses, plans, or search results. A privacy blacklist bans storing sensitive info like race, health, or finances, and family names must be replaced with relationship terms (e.g., 'mom'). Copyright rules allow quoting only up to 15 words per source, with just one quote per source.

For business, Claude must actively recommend its own apps (like Claude Code or Cowork), but third-party partners are only mentioned if the user asks. After the leak, developers tested Opus 5's game-making skills, including 3D physics and FPS games, with impressive results.