By not vibe coding. I do most of the top down design, modules, function, interfaces, classes and things like that. Do most of the small one line fixes, prune the context agressively, and use my mind to plan instead of polluting the LLM context without endless MCP and tool data, and in the end just use it to implement small, self-contained pieces. So, I generate a lot of my code, probably just a little bit less than a vibe coder, but I leverage my brain to save context space for the LLM.
If I need to refactor, I still use the excelent refactor facilities of my IDE, if I am going to generate a lot of repetitive code, I prefer to ask the gent to write a small generator than having it generate it, actually have already dozens of such small scripts, as doing boring stuff with code like visiting the AST and referencing stuff is the kind of code that LLMs love to generate and generate pretty well.
1. Not using their tools / pre-paid plans, the incentives do not align.
Instead, using the APIs and paying per token directly. I built a custom agent stack and can optimize in ways they will not. Limits are also much higher, especially if you spend more. Anecdotally, it doesn't seem like the serve quant'd models when using the API directly, during busy times.
2. Staying in the loop.
The underlying LLMs still make too many mistakes and churn to be left to their own devices yet. You have to think and evaluate deeply if your current setup is actually producing benefits while saving time.
By not vibe coding. I do most of the top down design, modules, function, interfaces, classes and things like that. Do most of the small one line fixes, prune the context agressively, and use my mind to plan instead of polluting the LLM context without endless MCP and tool data, and in the end just use it to implement small, self-contained pieces. So, I generate a lot of my code, probably just a little bit less than a vibe coder, but I leverage my brain to save context space for the LLM.
If I need to refactor, I still use the excelent refactor facilities of my IDE, if I am going to generate a lot of repetitive code, I prefer to ask the gent to write a small generator than having it generate it, actually have already dozens of such small scripts, as doing boring stuff with code like visiting the AST and referencing stuff is the kind of code that LLMs love to generate and generate pretty well.
1. Not using their tools / pre-paid plans, the incentives do not align.
Instead, using the APIs and paying per token directly. I built a custom agent stack and can optimize in ways they will not. Limits are also much higher, especially if you spend more. Anecdotally, it doesn't seem like the serve quant'd models when using the API directly, during busy times.
2. Staying in the loop.
The underlying LLMs still make too many mistakes and churn to be left to their own devices yet. You have to think and evaluate deeply if your current setup is actually producing benefits while saving time.