The conclusion in the article summarizes my experience of using LLMs for professional software development for past 2+ years.
I am an advanced programmer. I used ChatGPT standalone and Cline with Gemini Flash in VS Code (with code/markdown context). The application scope is full SDLC - domain exploration, requirements, modeling, tdd, ci pipelines, delivery (docker/swarm/k8). The coding stack is TypeScript, mostly backends past years.
As for coding specifically, I do not rely on LLM for scaffolding and framework things. They deliver <5% of code in a real project and always 100% better to copy/paste from earlier projects than to rely on LLM's probabilistic code let alone rarely accepted code quality.
I definitely do not rely on LLM coding me domain things, as whenever I tried, the result from LLM was on a final account, 2-3 times longer to get to the acceptable (!) level compared to my own 1 time long top-notch (!) result.
Apart of the quality and the time spent, its "wall of text/code" is annoying and distracting by irrelevant solutions, conclusions, notes etc. Despite my all-time context is "remember to give me super-concise answers".
It is like OOP vs procedural, UML vs no-diagramming, you have to know exactly what TO DO with them (LLM too), otherwise you are buried exploring a pile of WHAT NOT DO with them slop for ever.
