# How Anthropic Uses Claude Itself: The Complete Prompt for Plan Mode Was Leaked

The core of hundreds of words in Plan Mode isn't about making AI do more—it's about preventing it from acting.

- After entering Plan Mode, modifying code is completely prohibited. Only read, search, and ask. The prompt is hard-coded with "this requirement takes priority over any other instructions"—physically locking down write permissions.

- Maximum 3 agents running simultaneously from different angles to propose solutions. For new features, one examines simplicity, one examines performance, one examines maintenance costs. Let AI debate with itself first, then give you the choice.

- Plan submissions can only use dedicated tools; prohibited from asking "is this okay?" in the conversation. AI is too good at bypassing approvals using natural language—directly block it.

The most expensive part of writing code with AI isn't tokens—it's rework. Let it read and understand first before taking action. It looks slow, but it's actually fastest.