The Claude Code agent can be configured to operate in a neutral state using the --safe-mode flag. This mode ignores custom instructions, skills, plugins, hooks, and MCP servers, allowing the AI to function without user-specific bias.
The developer created a custom skill, /ccu, to automate this process. It runs a prompt in a clean instance by creating a temporary directory with empty work and home folders, ensuring no configuration files or history from the main environment leak into the session.
The tool uses the existing OAuth token from the macOS Keychain to maintain the user's subscription status while keeping the profile data isolated. Permissions for specific tools are managed via a local settings.json file, and the temporary directory is marked as trusted to bypass security prompts.
The /ccu skill generates a transcript.md file from the raw session log for the user to review. This approach allows for quick, unbiased checks by offloading tasks to a clean instance without requiring a full manual setup.
