Something about AI-based code generation seems lazy or sloppy. After all, "vibe-coding" is often used to convey poor quality. And giving an AI agent short, poorly-reasoned prompts tends to give at best inconsistent results.
Recently I started experimenting with writing detailed specification files (in Markdown) and having an LLM generate the required code. I have found using a single function, like parse_settings or parse_prompt, along with a known struct type that is returned is a good strategy.
Some notes:
• I can get reliable output from Qwen 3.8 27B (actually Bonsai Ternary 2).
• Code generation takes about half an hour even for short 150-line outputs.
• I can switch models (for example, using a frontier model in the cloud) and the specifications can be reused.
• I can just maintain the specifications instead of the actual code, which could be simpler.
• It is possible to use an LLM to "debug" a specification and resolve any inconsistencies, which is an important step in this approach.