Established local AI coding environments using Ollama, OpenCode, and Claude Code. Evaluated models including Qwen, Llama, and DeepSeek for code quality, instruction-following, and hardware efficiency. Developed structured prompting and context-management workflows for multi-stage development coordination.
Methodology
The workflow was developed through iterative experimentation: define a task, generate code with different models and prompting strategies, evaluate the output, refine the approach.
Key Findings
- DeepSeek Coder 6.7B — best for focused code generation
- Qwen2.5-Coder 7B — best for constrained generation with strict conventions
- Llama 3.x 8B — best for planning and critique
Structured Prompting
The most important discovery: prompting methodology matters more than model selection. The multi-stage workflow (Plan → Scaffold → Implement → Audit) consistently outperformed single-shot generation regardless of the model used.