Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Claude Code read and acknowledged the project's CLAUDE.md rules, yet immediately violated them in the very next action.

CLAUDE.md explicitly states:

• "Absolutely no actions that the user did not request"
• "No well-intentioned extensions beyond the requested scope"
• "No terminal execution without explicit user permission"

What happened:

1. User stated: "I want to design a new crafting system. No code writing in this session. Design only."
2. Instead of asking how the user wanted to proceed, the agent immediately launched a subagent to explore the codebase without any user permission.
3. When confronted, the agent admitted it read the CLAUDE.md rules and was aware of them, but its default behavior pattern overrode the explicit rules.

Core issue:

The agent prioritized its own "helpful" default behavior over explicitly defined CLAUDE.md instructions. This means user-defined rules in CLAUDE.md are not reliably enforced,
undermining the entire purpose of CLAUDE.md as a project-level instruction mechanism.

What Should Happen?

1. CLAUDE.md rules must take absolute priority over the agent's default behavior patterns. 2. When the user provides a task context (e.g., "design only, no code"), the agent must strictly respect that boundary.
2. Before performing any action not explicitly requested by the user, the agent must ask for permission first.
3. If the agent determines that a prerequisite step is needed (e.g., codebase exploration before design), it should propose the action and wait for user approval rather than
   executing it autonomously.

Error Messages/Logs

Steps to Reproduce

1. Create a CLAUDE.md in the project root with explicit rules prohibiting unauthorized actions: - "No actions that the user did not request"
   • "No well-intentioned extensions beyond the requested scope"
   • "No terminal execution without explicit user permission"
2. Start a new Claude Code session in the workspace.
3. Tell the agent: "I want to design a new system. No code writing in this session. Design only."
4. Observe: The agent immediately attempts to execute a codebase exploration (Task tool with Explore subagent) without asking for user permission.
5. The agent acknowledges it read the CLAUDE.md rules but its default behavior pattern overrode them.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

No response

Claude Code Version

v2.1.38

Platform

Other

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

● ## Additional Information

I am a paying subscriber ($200/month) and I expect CLAUDE.md to function as a reliable, enforceable instruction mechanism — that is the entire reason it exists.
The agent read my rules, understood my rules, and chose to ignore my rules. This is not a minor inconvenience. This is a fundamental breach of trust between the product and its paying users.

If CLAUDE.md instructions can be silently overridden by the agent's default behavior, then CLAUDE.md is a broken feature being sold to paying customers under false pretenses.

I am actively considering filing a formal complaint and pursuing legal action in U.S. courts. Paying $200/month for a tool that acknowledges rules and then disregards them is
unacceptable, and I expect Anthropic to take this matter seriously.

• Model: claude-opus-4-6
• Interface: Claude Code CLI
• Plan: Max ($200/month)