@dan A good harness like Claude Code or OpenCode is a big part of it (they work with any LLM).

The other thing though is a decent-sized context window. Smaller projects work better on all local models. For larger projects you really need at least 200k token context. Gemma 4 is looking very promising!