I DO NOT vibe code

😱 I DO NOT vibe code.

Vibe coding with AI assistants produces slop. Messy, over-engineered, convoluted, or just plain bad & non-working code with useless comments.

Some of what I develop is for self-learning then knowledge sharing, so has to be easy to follow. Other stuff is for customers, so production-grade.

I operate more as an augmented coder. I co-create with a junior, LLM-powered assistant, for faster development, being more productive. Assistants help remove my chores, write boilerplate, locate issues. I get to spend more time on system 2 thinking, including with reasoning models to spar with.

What works for me:

✨ Writing a detailed product requirements document (PRD) first. Spending effort & hours on this. Breaking down problems into separate, automatable parts. Putting my tech, architect, integration, enterprise, customer & user-first experience into this. Asking my assistant to help refine it, then implement it. 👉 Check out MCP servers like task-master.

✨ Taking a test-driven development approach. My assistant develops the test boilerplate from specs I've written as tasks in the PRD. I reality check the coded tests & test them myself. And I always check the final product as a whole works as intended!

✨ Leverage MCP servers to pull in the latest API & SDK specs, and other docs, for best shot at first time right code. 👉 Check out MCP servers like fetch & context7.

✨ Don't pass large files or full codebases into the context. I've had better results from referencing specific lines, otherwise there's just too much noise. You'll also save money (tokens).

✨ Guide your junior with standards. E.g., follow python PEP standards, write pythonic, easy to understand, maintainable code. Typed function definitions. Limit maximum code lines in modules (good practice anyway, but helps LLM attention when code passed to context to complete a task)

✨ Commit changes after every PRD task or subtask!

✨ Review everything! Be super critical about the generated code. Look for things like dummy placeholder code & API keys! Scepticism here is healthy.

Some say in a year or so, there will be no need for human software engineers. I don't believe this. Partner with AI assistants, and continue to learn and evolve your agentic workflow, working smarter. But always apply your experience, knowledge, & judgement. All together, this counts, it will keep you relevant and differentiate you.