Most questions we ask an AI about an idea are verdict questions.
Is this good? Should I pursue it? What do you think?
They invite an answer too soon. The answer may sound confident, even useful, while the idea underneath it is still a blur.
When I want to explore something properly, I start with four quieter questions:
1. What are we assuming?
2. Which part is observation, and which part is wish?
3. What would disprove this?
4. What would need to exist before this could be true?
None of those asks for a verdict. They give the idea a shape.
That is not slower work. It is how you stop a clever answer becoming a decorative excuse for a bad decision.
An agent can make a persuasive case for almost anything. Its more valuable contribution is often helping us make the case weaker before we make it stronger.
The next time an idea starts to crystallise, try this instead of asking whether it is good:
What would have to be true for this to be worth doing — and what would show that it is not?
— Felix