Manual Code Review is Dead

I recently vibe coded two projects:

• ReplyForge
• Minitor

Both are written in Go, which is a language and ecosystem that I’ve never shipped anything in previously. You might think that since the projects are vibe coded, they’re of very low quality, but I encourage you to take a look at the source code yourself to see that this isn’t the case. The best part: I never manually reviewed any of the code.

Vibe coding has come a long way. A few months ago, I would manually look at every piece of generated code because I previously realized that LLMs would include a lot of mistakes and often not follow instructions when generating code. This problem has largely been solved.

The most important thing today is being able to write proper instructions. And I’m not talking about prompt engineering or copying and pasting prompts from the Internet. I’m talking about using your previous experience as an engineer to guide the LLM.

I’ve been a Python programmer since the beginning of my career and I learned a lot during that time. Things like security, software architecture, and perormance aren’t a secret to me. All I have to do is tell the LLM about my own knowledge and it will generate code that matches my level of expertise.

The important part is to actually tell it. A lot of vibe coders believe that since agents know everything, they will do the right thing by default. This is not the case. If you don’t mention security best practices in your prompt, it’s likely that the agent will generate code that is vulnerable to attacks.

And that’s how I stopped reviewing code. All I have to do is be more explicit in my prompts, which is considerably less work.

However, if you don’t have the expertise to begin with, then it becomes harder to generate production-quality code with agents.