If found it better to split in smaller tasks from a first overall analysis and make it do only that subtask and make it give me the next prompt once finished (or feed that to a system of agents). There is a real threshold from where quality would be lost.