This site only uses technical cookies required for it to work: no tracking, no profiling. Cookie Policy

Skip to content
All terms

What is vibe coding?

Building software by describing what you want to an AI and accepting the code without really reading it. Fast for prototypes, risky in production.

Vibe coding is software development where you describe the desired result to an AI and accept the generated code without really reading it, iterating prompt by prompt until it "works". The term is Andrej Karpathy's (early 2025) and was named word of the year by Collins Dictionary in late 2025: meanwhile the practice went mainstream, with a large share of new code now written with AI assistance. For prototypes and experiments it is liberating, validating an idea in hours instead of weeks even for people who do not code for a living, but the risk starts once that code reaches production without ever passing through expert hands: nobody in the company really understands it, tests are missing and dependencies were picked at random. That scenario is where the term "vibe slop" comes from, low-quality AI code accepted unread together with the production incidents that follow from it.

The good side and the dark side

For prototypes, experiments and internal tools, vibe coding is liberating: ideas validated in hours instead of weeks, including by people who do not code for a living. The problem starts when the prototype reaches production without ever passing through expert hands: code nobody in the company understands, no tests, random dependencies, security holes. The phenomenon already has a name, "vibe slop": low-quality AI code accepted unread, and the production incidents that follow from it.

The practical rule

You do not ban vibe coding (the productivity gains are real), you scope it: free rein on prototypes, while anything going to production requires competent human review, tests and clear ownership. And for vibe-coded software that already became critical without anyone noticing, the path is an honest assessment and a cleanup: exactly the use case of my Software Rescue service, which treats ungovernable AI code as the new legacy. The bill is not paid in code alone, though: when nobody on the team can explain how what they shipped actually works, that is cognitive debt, and no rewrite pays it off.

  • AI Agents (Agentic AI) · AI systems that go beyond answering: they plan, use tools and take actions autonomously inside your processes.
  • Forward Deployed Engineer (FDE) · An engineer working inside the client's company, side by side with its teams, accountable for the outcome, not the slides.
  • Shadow AI · The use of AI tools at work without approval or oversight: employees pasting company data into ChatGPT and the like.
  • AI slop · AI-generated content produced at volume without care: mediocre text, images and code polluting the web, inboxes and repositories.
  • Claude Code · Anthropic's tool for agentic software development from the command line, an agent that lives in the terminal rather than an editor.

A term that hits close to home? Let's talk.

CONTACT ME