The 80:20 pattern
Across every serious system I've built through conversation with AI, the same shape appears. Roughly eighty per cent of the demonstrable functionality arrives fast: features that would once have taken days or weeks reach functional states in hours. Then the last twenty per cent turns out to be a different kind of work entirely. Security review. Accessibility. Edge case handling. Permission systems. Deployment configuration. The journey to a convincing prototype is fast; the journey to something you could responsibly put in front of other people is not.
That last twenty per cent is not a residue of polish. It's where the expertise the tools were supposed to replace turns out to still be load-bearing.
What it looks like in practice
Vibe-coded systems tend to emerge beautifully typed but incompletely wired: components rendering mock data, handlers that exist but do nothing, APIs that are defined but never called. On Artspace, a full-stack arts-sector platform built under a deliberate rule of writing no code by hand, the prototype came together in days while the permissions model, the part deciding who can see whose data, demanded exactly the architectural knowledge the conversation couldn't supply. On the Fantasy Premier League engine, twenty-plus verification passes confirmed the code was structurally correct while a semantic error (end-of-season player prices standing in for point-in-time prices) sailed through every one of them.
The failures cluster into a taxonomy, six modes from configuration blindness to regeneration cascades, covered in the six ways vibe-coding breaks.
What it isn't
It isn't an argument against the tools. The compression is real, and for non-technical creatives it's the difference between making software and not making it. The point is narrower: the tools don't eliminate the need for knowledge, they transform what knowledge is needed. The gap between demonstration and deployment is where that transformed knowledge lives.
What fills it
Not a computer science degree. The research suggests a small, learnable floor of concepts, Minimum Viable Literacy: five domains that let a non-technical maker cross the last twenty per cent with judgement rather than luck.
Where the term comes from
It emerged from instrumented case-study work in 2025, two CDT industry challenges built and documented end to end, and is being tested in participant studies through 2026. The fuller story is in what happens when you let AI write all your code and on the research page.