> You need to fact check it (especially with LLMs)

That's true... but with every day that passes it's increasingly less true. Today's best high-end LLMs are highly optimized towards fact checking themselves and each other. The winning move is to get a report from Claude and have both Kimi K3 and GPT-5.6-Pro adverserially review it. Then synthesize a fixed version with any one of them. Then adverserially review it again in incognito mode.

This sounds like a lot of work, but it takes literally two minutes of effort and it works amazingly well.