What model were you using? This is true for e.g. the best Qwen model, which I believe to be too dumb to be useful, but not GPT-5.0 or above in my experience (and I consider myself to be a pretty demanding customer).