Armin Ronacher shares his experience with Claude's features, particularly focusing on YOLO mode and plan mode. He initially favored YOLO mode for its flexibility but found plan mode cumbersome due to its frequent requests for approval. Ronacher's alternative approach involves collaborative iteration with the AI, generating markdown files through a question-and-answer process. However, after discussions with others who prefer plan mode, he sought to better understand its functionality by examining Claude Code's implementation. In Claude Code, a plan is represented as a markdown file created within a specific folder. While it lacks complex structure, the main difference lies in the agent's behavior when in plan mode. The agent operates under a read-only directive, with specific prompts guiding its actions. Notably, the agent edits its own plan file, which is not something that can be easily replicated outside of plan mode. Ronacher outlines the phases of plan mode, from understanding the user's request to finalizing the plan, emphasizing that the process is mostly about reinforcing prompts rather than altering the agent's capabilities. Ronacher concludes that while the differences between using plan mode and manually crafting prompts may not seem significant, there are nuances in the user experience. The hidden markdown file offers a level of organization, but the custom prompts that guide the agent’s behavior in plan mode create a distinct workflow. He notes that entering and exiting plan mode involves a straightforward prompt that changes the agent’s operational context, reinforcing that the tool’s function is primarily about establishing a structured approach to user requests.