Minimum Viable Literacy.

The smallest set of technical concepts a non-technical maker needs to direct, debug, and judge AI-built software. Not a coding curriculum: the floor below which shipping becomes luck. A framework from my PhD research; this page is its canonical definition.

The five components

01

Systems thinking

Users, goals, end-to-end flows. The foundational one: without it, prompts drift into disconnected features that never cohere into a usable system.

02

Basic architecture

Not implementation depth. Practical distinctions: frontend and backend, client-side and persistent storage, user roles and permissions, public and private data. Enough to ask the right questions.

03

Configuration literacy

Knowing that external services need configuring, that credentials and provisioning exist, and that generated code will fail without setup the AI cannot do for you.

04

Error literacy

Reading logs, naming the failing component, describing the gap between expected and actual behaviour, and asking for targeted fixes rather than wholesale regeneration.

05

Safety literacy

Knowing where deployment risk concentrates (authentication, storage rules, data boundaries, credentials), where the model is blind to dangers it cannot perceive, and when to stop and get expert review.

Where it comes from

Each component maps to failures I watched happen, not to a syllabus. Across two industry-scale case studies, every documented breakdown traced back to one of these five being missing. The framework is what's left when you ask, for each failure, the smallest piece of understanding that would have caught it.

The failure modes it catches

Failure modeWhat it looks likeCaught by
Configuration blindnessGenerated code assumes infrastructure that was never provisionedConfiguration
Permissions mismatchFunctional features with broken access boundaries at the edge casesArchitecture
Data boundary violationsSystems lose track of what data belongs where, or to whomSafety, architecture
Dependency driftA change to one component quietly breaks anotherSystems thinking
Overconfident unsafe defaultsDangerous configurations accepted without warningSafety
Regeneration problemsBroad fix requests rebuild whole components, trading old bugs for newError

The six modes get their own treatment in the six ways vibe-coding breaks.

What it deliberately is not

It is not "learn to code, but less". None of the five components involves writing syntax. They're judgement capabilities: knowing what to ask, what to check, and when to stop. The depth required is also adaptive rather than fixed; a tool you build for yourself and a multi-user deployment with real people's data don't need the same floor.

Status

Version one of the framework comes from the case-study work and is written up in a draft paper. Two participant studies through 2026 test whether it holds for practitioners who aren't me: single-session first encounters, then 8–10 weeks of learning trajectories. The framework sits inside the wider question of the compression gap; the full programme is on the research page.

Want your team to have this floor?