Vibe coding outsources judgment along with execution
The term describes using AI tools irresponsibly — accepting whatever is produced without scrutiny or genuine accountability. Whether you’re vibe coding is not about which tools you use or how often. It is about whether you maintain understanding and direction throughout, or hand off the thinking along with the work.
Vibe coding
One-sentence definition: Vibe coding is working with AI in a way that outsources judgment along with execution — using the tools while abandoning accountability for the outputs.
The term was popularised by Andrej Karpathy and developed further by Simon Willison, who drew a useful distinction between vibe coding and vibe engineering: what experienced practitioners do when they use the same AI tools while remaining genuinely accountable for every output. The tools are identical. The difference is in what the practitioner maintains throughout.
What makes this different from disciplined AI use:
Disciplined AI use involves maintaining the ability to evaluate what an agent produces — to recognise when something is wrong rather than merely plausible, and to redirect with enough precision to improve it. Vibe coding can produce fluent, convincing output. What it can’t produce is output you can stand behind, because standing behind it requires the judgment that wasn’t maintained in producing it.
The failure mode is most visible when something goes wrong. A practitioner who has been vibe coding cannot say why the output falls short or how to improve it. They can accept or reject, but not direct.
What it looks like in practice:
- Accepting a generated literature review without checking whether sources say what the AI claims they say
- Delegating a task without a clear brief and accepting whatever comes back
- Producing writing that sounds like your thinking but doesn’t reflect it
What remains unclear
The line between disciplined delegation and vibe coding is not always sharp. For genuinely routine tasks — reformatting, reorganising, converting between structures — there may be little meaningful distinction between accepting output and confirming it is correct. The harder question is where that line sits for intellectually substantive work, and whether it shifts as models improve and practitioners develop better tools for evaluating what agents produce.
Sources
- Willison, S. (2025). Vibe engineering. simonwillison.net.
Notes
The contrast with context engineering is instructive: context engineering is the practice that makes disciplined AI use possible at scale. Vibe coding is what happens in its absence.