AI lies. The trick is making it double-check itself before it ships. I made a skill that forces AI to think more. Burns more tokens but improve accuracy. I suggest you use Meta CoT(Chain-of-Thought) https://arxiv.org/abs/2501.04682 to build a think skill to avoid this problem.

Thanks for the paper link! I'll check it out.