Can this be 'vibed'? No, but... we can get a long way if we keep a close eye on how the agent codes.
For the LLMs I'm using only local models. These models change on a weekly basis. Currently I'm using mostly MiMo-V2.5-Pro and GLM-5.1 but I’m also using Qwen-3.6-27B as a general all round workhorse
I'm finding that the AI is helping when it knows what I don't know, however as the project is getting larger, it is having issues with reusing existing code as I cant fit it all in the context. - I now have a working solution to this using a Knowledge Graph that has full repo understanding of the code
The best prompting technique seems to be to ask it for 3 options on how to do something, then I pick the one I'm happy with and tell it to "go ahead and code option N", then when it is done, I 'git diff' or use vim to view the changes.
This approach hasnt changed for this and other projects I’m doing