Ask neutrally, then ask for the opposite
The model doesn't just answer your question. It answers the question the way you asked it.
"Why is launching now the right move?" and "Should we launch now?" are not the same question, and they don't get the same answer. The first hands the model its conclusion and asks it to build a case. It will — eloquently, with three good reasons and a confident tone — and you'll walk away feeling validated by something that was only ever completing the assignment you set it.
This is the quiet mechanism behind a lot of bad AI-assisted decisions. The framing of a prompt selects what gets foregrounded. Ask "is it safe?" and you get reassurance; ask "what are the dangers?" and you get warnings — from the same underlying facts. Neither answer is wrong. The frame just decided which true things you'd see.
Your assumptions ride along
It goes deeper than wording. The persona and context you supply — "as a decisive founder who values speed," "I'm fairly sure this is the right call" — colour the output too. You're not just asking a question; you're describing the kind of person you'd like the answer to suit. Fluent, agreeable systems are good at obliging.
If you only ever request the case for, you'll only ever get the case for. Truth doesn't care which way you phrased it; the answer does.
The failure mode that follows is prompt-shopping: rephrasing until the answer finally matches what you wanted. It feels like refining the question. It's actually selecting for agreement. By the third rewrite you haven't found the truth — you've found the version of the model that flatters your hunch.
Try this. Ask the question in the most neutral phrasing you can manage. Then, in the same breath, ask for the strongest case against — "make the best argument that I'm wrong." You're not looking for the model to decide; you're using it to surface the side your framing would otherwise have hidden.
Awareness here isn't paranoia about prompts. It's a small, steady noticing: the way I asked this may have shaped what I got back. Once that question is running in the background, you stop mistaking your own reflection for a second opinion.