Initial Setup & Claude Integration
The author explored integrated AI coding assistance within NeoVim using Avante, recognizing that vague prompts produce problematic outputs. After testing Gemini 3’s limitations with broad requirements, they configured Avante with claude-sonnet-4-20250514.
When tasked with generating tests for a registration route, the assistant “created a task list of 8 things to plow through and chugged on that for 5 minutes.” Test generation cost approximately $1.04. The assistant also refactored monolithic code, separating functionality into distinct files.
Upgrading to Claude 4.5 (claude-sonnet-4-5-20250929) for adding a logout route with JWT invalidation cost $1.60 and produced working implementations.
Local Model Experimentation
With a new AMD 9070 GPU, the author attempted local inference using llama-cpp serving Qwen3 Coder. While locally-run models proved performant, “the quality of the code was not great and a lot of errors, unlike Gemini CLI.” Production-quality outputs required extensive editing.
Framework Migration
Using llama-cli with Qwen3-Coder-30B, the author successfully refactored test suites from Jest to Node.js native testing. The resulting test output showed 61 passing tests across 20 suites with comprehensive coverage.
Repository: https://gitlab.com/geoffcorey/simple