Raw LLM Responses

I had a similar thought to you until I learned how to use it more effectively, that’s where the real alpha is. It’s all about context and planning. You have to create extremely detailed documentation (which you can use AI to do). There is some prompt engineering that goes into it as well. For example, you don’t just say “make me a full stack to do list web app” and expect it to one shot it first try. Ideally, you research the best tech stack, create a PRD, create a implementation plan, and keep track of changes made as you go along. Utilize an IDE like cursor or windsurf to give these documents and your code base as context to the models and you will get much better results. Also it’s important to frequently push to git and revert anytime you run into errors. A big mistake I see people making is having the LLM debug which usually just results in more errors and slop. Instead, revert and improve your prompt with better context until it gets it right. This is really high level overview but you can find really good guides on YouTube and X. Also, I’ve found o1 or grok 3 thinking best for creating documentation and Claude sonnet 3.5 or 3.7 best for doing the actual coding. Careful using 3.7 though as it likes to go crazy and implement a lot of stuff you didn’t ask for so requires good prompt engineering to keep under control. I personally have been able to successfully create full stack production apps and deploy them but there was of course a learning curve and it took a few tries.