I always ask my AI to be contrarian and honest. My usual process to get there is to brainstorm back and forth with the AI, until I’m happy and we jump together to plan mode.
However, most models are over-eager to please and they propose way too much in an open brainstorming session. A better way that I found to try to reach the shared understanding without increasing the scope too creatively was Matt Pocock’s “grill me” trick that he also ships as a skill.
“Grill me” does essentially what it says and it forces the LLM to explore all the potential branches in a way that is much more structured.
After being “grilled”, I ask the assistant to write the conclusions of our session to a file, and then I edit that file directly by hand (like a barbarian!) to consolidate, course-correct and reason through the main code modules. I don’t call plan mode anymore — this file becomes the plan.