Simon Willison and the Boundary of Vibe Coding
Simon Willison became one of the clearest contemporary interpreters of vibe coding by insisting that the term should describe a specific practice—letting AI generate software without closely reviewing the code—not every use of an LLM by a programmer.
Historical terms need boundaries if they are going to explain anything
Once “vibe coding” became popular in early 2025, it quickly began to mean different things to different people. Some used it for any AI-assisted programming; others used it specifically for building software without paying much attention to the generated source.
Simon Willison became an important interpreter because he documented the original Karpathy description and repeatedly argued for preserving its narrower meaning.[1]
Willison preserved Karpathy’s original February 2025 wording
On February 6, Willison published a quotation of Karpathy’s post describing a workflow of talking to Cursor Composer, accepting changes wholesale, copying errors back into the model and allowing the codebase to grow beyond easy comprehension.[1]
That archival act matters because later paraphrases often omit the most distinctive part of the definition: the deliberate lack of code review.
The original term was playful and self-aware
Karpathy presented the workflow as amusing and especially suited to throwaway weekend projects rather than as a universal professional-development doctrine.
The phrase described a relationship to generated code
What makes the original practice distinctive is not merely that an LLM writes code, but that the human intentionally stops treating line-by-line understanding as the center of the activity.
Willison explicitly separated vibe coding from responsible AI-assisted programming
In March 2025, Willison argued that using an LLM to help write code does not automatically count as vibe coding. If the developer reviews the code, tests it and remains able to explain what it does, he considers that ordinary software development using AI assistance.[2]
This distinction is valuable for CodeHistory because it prevents a single catchy term from swallowing the entire field.
His boundary is accountability, not anti-AI skepticism
Willison is an active user of language models for programming and has written extensively about productive LLM-assisted workflows. His objection is not to AI-generated code; it is to confusing an intentionally low-oversight style with accountable engineering.[2]
Production software carries obligations beyond whether a demo works
Security, accessibility, maintenance, performance and cost remain engineering concerns even if AI writes most of the implementation.
Understanding can be delegated only up to a point
A developer responsible for production behavior needs enough understanding to investigate failures and make informed trade-offs. Vibe coding deliberately relaxes that expectation.
Willison also defended vibe coding as a useful low-stakes practice
His March article is not a dismissal. He argues that vibe coding can help beginners create custom tools and can help experienced developers build intuition for what LLMs can and cannot do.[2]
This balanced framing helps explain why the practice could be both empowering and risky.
Semantic diffusion became part of the story within weeks
By May 2025, Willison was criticizing books and commentary that used “vibe coding” to describe professional production-grade AI-assisted development.[3]
The speed of that shift is historically interesting: a term could move from one person’s precise playful definition to an industry-wide marketing label in under three months.
Willison later proposed new language for the accountable end of the spectrum
In October 2025, he suggested “vibe engineering” for experienced professionals who use LLMs aggressively while remaining responsible for the resulting software.[4] In a later update he noted that “agentic engineering” appeared to be gaining traction instead.
This shows terminology evolving in response to a real need: distinguish low-oversight creation from rigorous model-assisted engineering.
Why Simon Willison belongs in a pioneer category even though he did not invent the tools
Histories of technology need interpreters as well as inventors. Willison documented the original coinage, tested the practice publicly and supplied a vocabulary for its limits.
Contemporary reporting also relied on the same Karpathy definition when explaining the phenomenon to broader audiences.[5] His role is therefore definitional and historiographic: he helped preserve what the term originally meant while the industry rapidly expanded it.
Works Cited
- 01Simon Willison — A quote from Andrej Karpathy (Feb. 6, 2025) simonwillison.net
- 02
- 03
- 04Simon Willison — Vibe engineering (Oct. 7, 2025) simonwillison.net
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead