vibe coding: a collaborative autoethnography

the question

vibe coding — prompting a model in natural language to write code, with relatively little review of what comes back — went from a joke to a default workflow extremely fast. the discourse about it is polarised and mostly speculative: either the democratisation of programming, or the end of learning to program.

neither claim rests on much evidence about the actual experience of doing it. so rather than survey people about it, we set up a study where the researchers were the ones doing it, documenting as they went.

why collaborative autoethnography

chang and colleagues define it as a method where researchers work in community to collect their own autobiographical material and then analyse and interpret it collectively. in practice that means individual written reflection followed by group discussion, over and over.

i chose it for two reasons. first, i wanted the participants to have as much say as possible not only in what data got collected but in how it was analysed and interpreted — and these are students whose own curriculum has already taught them empirical research practice, so they could genuinely do that work. second, the method forces shared meaning-making rather than a researcher deciding alone what a reflection meant.

how it ran

• nine weeks, six undergraduate co-researchers at a us university, each building a project they chose themselves. irb-reviewed and determined exempt.
• weekly written reflection as the collection instrument, capturing not just what got built but where people stalled, what they accepted without understanding, and how it felt to call the result theirs.
• whatever tools they liked — the drift between tools over the nine weeks became data in its own right.
• collaborative analysis, with the co-researchers coding the corpus and settling disagreements among themselves rather than deferring to a facilitator. that was the methodological commitment, not a shortcut.

the point of the method is that the people doing the thing get to say what it meant.

this paper is currently under review, so i’m keeping the findings off this page entirely. happy to talk about them — get in touch.

why this sits in my research

this looks like a departure from grades until you notice it’s the same shape: a person navigating an opaque system that produces consequential output they’re expected to evaluate without full visibility into how it got there.

it’s also my most direct look at how people build literacy around ai tools in contexts that matter to them — the thread i’m carrying into my current work on ai-run hiring interviews, and a close cousin of the measurement literacy work.