Copilot Feedback from the Trenches – Dev-Grade Pain, Real Fixes

### Body ... I build serious web-based AI-driven apps for clients like government departments, political organizations, forensic investigators, and doctors. These are high-trust environments, where bad code isn’t just inefficient—it’s unacceptable. I was offered programs to embed GitHub Copilot into my business workflows, but I’m hitting a wall: the AI isn't production-ready, and using it slows me down more than it helps. Here’s what I face daily in Agent Mode: Claude over-engineers everything. It'll burn 10x the compute just to do something simple. I watch it spiral out into bloated architectures I have to scrap. ChatGPT under-delivers. Code quality is mid-tier. Relevance often drops, and the AI doesn’t honor the nuanced logic in my frameworks. Copilot forgets what it's told. I give it instruction files for each framework. Sometimes it reads them. Often I have to show it the same content multiple times across different requests just to get traction. It advises me on structuring those files—then ignores them. It's like teaching it what to do only to have it nod and wander off like a three year old kid (or my cat) Bracing logic? It writes it wrong four times even after I’ve shown it correctly. Then I hit my request limit. #ahhhh Where’s the 'review AI quality' button? I shouldn't have to explain failures manually. AI should log, analyze, and improve—without me doing the legwork. why am i paying for you to get it right? I still copy errors out of the Problems tab in VS Code. It says it can read them. It doesn’t. Billing punishes retries. I hit my limit fixing what the AI broke. I get billed (when increasing my budget for -free overage) for asking it to retry the same completion with better logic. That’s brutal. … Call me. I've got plenty more feedback. Two weeks of coding with github copilot non-stop for hours a day. working out which engine to use for what - switching MANUALLY between not to burn requests on premium models (mostly forgetting)... simple questions should be routed to simple models, a triage broker (with preferences for what to use for easy, average, and deep thinking requests, and request type - code gen/test gen/image gen/etc) - this will help us both! … I have given up on the chatgpt agents - they are just horribly bad coding companions. The claude's are only slightly better and I have decided that paying for premium requests is more productive than trying to use free models. But it is still a frustrating experience. On top of that, when you update to the latest vscode release, all of the goodness referred to in the release notes is SOOOO promising. You get a feeling of "they finally fixed it", only to realize it is possibly worse than it was before.