In the first part I talked about, vibe coding being better with tool usage, than actually writing code. Also to always correct the output is somewhat hard from a terminal, not being a full editor. Much background is not shown and, you probably cannot know it in the first place, for vibe coding.

But there is a tool for doing just that: Testing/TDD. Instead of always inform the LLM about a task for ever and ever I let it write some tests first. That do the correcting step for me. But is the tool good enough to write tests that, are correct, inform the network about the task? Will the resulting bug fixes, that the LLM than writes, be actually good? Will we be faster now?

Due to the additional indirection, you see even less code and are even more rely on the hope, that the LLM does not hallucinate.

Strikingly the Code gets overengineered a lot. Like write teen times the amount of code needed (number not made up). That extra nine need have written tests, need to be debugged, need to be understood and then thrown away. It not only costs money/credits it also costs a lot of time. The one fold that was need, was thankfully not but in the example, but still important parts where misunderstood and rewritting and debugging with the Pycharm Debugger was easier, than telling what was wrong. Because these corrections also would have been partly false.

What would I still do with LLM

Tool usage: debugging with linter and type checker
(Technical) Documentation
Partially checking the documentations of used libraries when not sure, what to look for.
Get ideas how to solve things, I am not so familiar with yet. Then check these manually for problem and loopholes.
Rewrite code to other code stiles, without changing what the code actually does.
Use it for smaller, sites jobs. I can check the results later on.

So in the end, after trying out vide coding my usage of LLMs has changed. Still I prefer similar to before. Write/Design the code myself. But it can speed up at the ruff edges of the code. Like when the test code needs a tiny change in the function under test to load the test case. The speed-up in debugging is there, it helps.

Last modified on 2025-07-28