I have not used it to write big software, mostly tools or POCs that I was interested in. Crafting properly worded instructions worked wonders. In one case, I told my local model to analyze a given API and look for a given language's libraries. Then incorporate those to do what I needed to do with the APIs. The added bonus was that I could tell it to find the correct version of the API and generate code against a small subsection of it.
None of this are for my work/job and if LLM didn't exist, these tools simply would not exist.