Interesting phenomenon: those code generation tools always underestimate their own capabilities.

Look at the planning schemes they provide—things they can definitely do are arranged as if paying in installments. First and second weeks do this, third and fourth weeks do that. And what’s the result? If you follow their approach, in less than 2 hours, you can complete the workload planned for these four weeks.

This phenomenon of overestimating work cycles and underestimating execution efficiency is indeed quite funny. What does it imply? There is a real mismatch between code generation ability and task planning capability. Either the system is intentionally being conservative (lowering expectations), or the performance gap between the planning module and the execution module is much larger than expected.