I'm not sure what to make of these technologies. I read about people doing all these things with them and it sounds impressive. Then when I use it, it feels like the tool produces junior level code unless I babysit it, then it really can produce what I want.
If I have to do all this babysitting, is it really saving me anything other than typing the code? It hasn't felt like it yet and if anything it's scary because I need to always read the code to make sure it's valid, and reading code is harder than writing it.
Because using AI takes some practice:
- you should document your best practices in a file and point it to the LLM (the standards are @claude or @agent markdown files
- you should manage context (the larger it gets the weaker the output)
- you should use good and clear prompts
- you should generally make it generate a plan with the requirements (business logic changes focused) and then follow and review the implementation plan (I generally produce both in two different markdown files).
- only then you let it code
The last phase, isn't even the most important to be honest, you can do it manually. But I have found that forcing myself through the first two and having AI find information in the codebase, edge cases in the business logic, propose different solutions, evaluate the impact of the changes is a huge productivity multiplier.
Very often I'm not worn out by the coding part, again, I can do it on my own, it's the finding information and connecting the dots the hard one. In that, it excels and I would struggle (mentally) to go back to jumping from file to file while keeping track of my findings in notes to figure out the wheres, whats and whys.
> and reading code is harder than writing it.
This is the things thar gets me the most. Code review is _hard_. So hard that I'm convinced my colleagues don't do it and just slap "LGTM" on everything.
We are trading "one writer, one reader" for "two readers", and it seems like a bad deal.
I'm always puzzled by these claims. I usually know exactly what I want my code to look like. Writing a prompt instead and waiting for the result to return takes me right out of the flow. Sure, I can try to prompt and ask for larger junks, but then I have to review and understand the generated output first. If this makes people 10x faster, they must have worked really slow before.