Why I stopped using LLMs to generate code (and what I use instead)
The author clarifies they are not anti-AI, using LLMs daily for tasks like parsing prose and summarizing documents. However, they stopped using LLMs for generating application code because the output, while visually impressive from tools like Bolt, Lovable, or v0, lacks critical production requirements. Common issues include missing or incorrect database schemas, authentication that stores tokens in localStorage (a security problem), no multi-tenancy (every query returns all users' data), mismatched OpenAPI specs, and absent migrations. The author argues this is structural: LLMs are stateless across the context window, leading to outputs that require 4–8 weeks of cleanup before the app can be shown to real users. They built an alternative but do not detail it in the excerpt.
LLM-generated code often requires weeks of fixes before production use, undermining developer productivity.