There's no universal solution. You pick what you like best and what works best for you. The main thing to keep in mind - you need to babysit the agent and give it guidance. You can't come up with a detailed 10-page document and tell it "go execute it" and step away for tea and come back to completed work. You can tell it "go complete step 1 of this document and test it". It will find mistakes, fix them. And then you can clear your context and tell it to go tackle step 2, etc.

Personal preference - even though Claude code does bloat context (initial prompt is very often 30K or more), I tend to get the best results from it. Agents focused on minimizing prompt (like Aider) don't work for me that well. But this also could be because at work I do use Claude or copilot, so I am used to that workflow.