AI can already build you a working interface on the fly. A new paper argues it's time to make sure that interface doesn't ignore decades of human-computer interaction research.
A researcher's framework, posted to arXiv on October 5, 2026, notes that chatbots like Claude and ChatGPT already generate functional user interfaces from plain-language requests, and proposes loading classic human-computer interaction (HCI) knowledge into that process as runtime "skills." That means encoding Nielsen's usability heuristics, Norman's affordance rules, Web Content Accessibility Guidelines (WCAG) criteria, cognitive-load limits, and mixed-initiative principles into machine-readable "skill.md" files the AI reads while building an interface. The paper calls the user-AI exchange a "Space to Think," treating interface generation as part of the user's thinking process rather than a bolted-on, finished product. The aim is to pack HCI expertise into inspectable, version-controlled files the HCI community owns and edits, instead of leaving UI quality to whatever a model absorbed during training.
That's a bet that usability and accessibility should be enforced inside the generation pipeline, not patched afterward by a designer reviewing output. It reframes WCAG compliance and basic usability checks as something closer to a linting rule than a design review: get the skill file right once, and every on-demand interface a model builds inherits the same baseline.
It's a tidy idea on paper. Whether a skill.md file can substitute for the judgment calls human designers make on messy, context-specific accessibility problems is the part no arXiv abstract can settle.