The fastest way to lose trust is to repeat something an AI told you and have it be wrong. One bad number in the wrong room and people stop believing the rest of you. It's the main reason people try these tools once and quit.
So I don't just trust the draft. I make the model attack its own work first.
Draft [N] points on [TOPIC]. Then identify [N] ways your own draft could be inaccurate, misleading, or missing critical context. For each, cite the specific claim and revise it.
Most of the time it flags its own shakiest claims, the stat it isn't sure about, the thing it implied but can't support. That's the gold. It tells you exactly where to aim your own fact-checking.
The guardrail matters here, so say it out loud: this is a focus tool, not a finish line. The model catches maybe 80% of its own weak spots. The other 20% is why you still read every number, date, and name yourself before any of it goes out. The self-check makes your verification faster. It does not replace it.