Have you tried Claude Code or OpenCode? I find working on defining a good issue / feature or even full spec, and then letting it run a plan based on that, and check it step by step rather a Ralph Wiggum loop. Then getting GitHub Copilot to do a review (I prefer Copilots reviews to Claude). Then get the agent to address to items. I’ve been using a custom voice app I built for voice integration into the terminal is good, when you can just explain more nuance that you just wouldn’t really type out can be good too (with screenshots). Lastly, creating a /reflect skill is cool, because after some time (often before a compact) you can ask it to reflect on your conversation to further refine the system prompt, skills etc. that would follow how you like things done your way next time first time.