Engineering Craft › AI-Assisted Development

Vibe Coding

Accepting AI code without reading it: fine for prototypes, risky in production.

Also known as: vibecoding, AI-generated prototype coding

Vibe coding is building software by describing what you want to an AI and accepting what it produces, with little or no reading of the code. You go by feel: run it, see if it looks right, ask for changes, repeat. The term was coined by Andrej Karpathy in early 2025.

It’s real and sometimes useful: people with no programming background build working tools, and experienced developers throw together prototypes in an afternoon.

Where it’s fine

• Throwaway prototypes and demos.
• Personal scripts and tools where only you are affected.
• Exploring an idea before committing to build it properly (MVP).
• Learning what’s possible, as long as you follow up by understanding it.

Where it’s risky

• Anything real users depend on: bugs, security holes and data loss hide in code nobody read.
• Security: injection flaws, exposed secrets, missing authorization checks (IDOR, secrets in Git).
• Maintenance. When it breaks, or you need to change it, no one understands it. The code is technical debt from day one.
• Hidden quality problems: duplicated logic, no tests, fragile behavior.
• Not learning. If you never read what you ship, you don’t grow (learning while using AI).
• Unaccountable code. You’re still responsible for what you deploy.

If you start that way

When a prototype is going to become a product:

1. Read it, properly (reviewing AI code).
2. Add tests, and check the edge cases.
3. Review for security: inputs, authentication, secrets.
4. Remove what you don’t understand, or learn it.
5. Restructure before building more on top.

Professional use of AI assistants is different: you stay the author, review each change, and keep the standards you’d apply to a teammate’s code (AI coding assistants).