Most AI workflow failures do not happen because the prompt is too short.

They happen because the prompt is the only thing holding the process together.

In long research tasks, especially business research, the model can start well and still drift later:

It summarizes before verifying.

It treats weak sources as if they were primary evidence.