After months of coding with LLMs, I'm going back to using my brain

a year ago

The author initially used LLMs (like Claude and Cursor) to build a new SaaS infrastructure in Go+Clickhouse, prioritizing speed over code cleanliness.
After weeks of development, inconsistencies and bugs emerged in the AI-generated code, leading to frustration and setbacks.
A code review revealed duplicated methods, inconsistent naming, and lack of structure, resembling work from uncoordinated junior developers.
The author shifted to a more hands-on approach, reducing reliance on AI for core coding and focusing on manual planning and debugging.
LLMs are now used for simpler tasks like renaming parameters or converting pseudocode, rather than writing entire codebases.
The author expresses concern about over-reliance on AI dulling problem-solving skills and the challenges non-coders face with AI-generated code.
Despite experimenting with various AI models and workflows, the author found limitations in handling complex queries (e.g., Clickhouse with large datasets).
The article critiques the hype around AI tools, noting inconsistent performance and the potential for "gaslighting" by benchmarks and influencers.
The author advocates for a balanced approach: leveraging AI as an assistant while maintaining human oversight and critical thinking.

Hasty Briefsbeta