The interface speaks for itself — or it doesn't work.
Don Norman's "The Design of Everyday Things" (1988) laid out a small set of design principles that became the foundation of HCI. The premise: a well-designed object tells you what it does, how to use it, and whether you got it right — without a manual.
The properties of an object that suggest how it can be used. A handle affords pulling, a button affords pushing, a knob affords turning. In software, affordances are mostly perceived — a button "looks clickable" because its shape echoes a physical one. When the affordance lies, the user fails.
Norman added "signifier" to the second edition because "affordance" had been overloaded. A signifier is the visible signal — a label, an underline, an arrow — that points to where action is possible. Affordance is the underlying capability; signifier is what the user can actually see.
The relationship between controls and their effects. Light switches that match the position of the bulbs they control; stove knobs aligned with their burners; volume sliders that go up when sound goes up. Bad mapping is why hotel showers and laundry machines are the running joke of UX.
Every action gets a visible, audible, or tactile response. Without feedback, the user is left guessing whether anything happened, and starts re-clicking, re-submitting, and double-paying. Spinners, toasts, ripple effects, haptic taps, sound — small cues, huge confidence.
Limits that prevent invalid actions before they happen. A USB plug that only fits one way, a date picker that won't let you pick before today, a tax form that grays out fields irrelevant to your filing status. Good constraints prevent errors invisibly; bad ones just frustrate.
The simplified picture of how the thing works that the user carries in their head. The folder metaphor on the desktop, the page metaphor in a browser, the chain of messages in chat. When the system's actual behavior diverges from the user's conceptual model, every action becomes a guess.
Whatever can be done should be visible. Hidden controls, ⌘-only shortcuts, and "if you knew where to look" features all violate this. The user shouldn't have to remember; the interface should show. Discoverability is what turns a powerful product into a usable one.
The same thing should always look the same and behave the same. Across screens, across products, and against the platform conventions users have already learned. Inconsistency forces users to constantly re-learn, undermining every other principle on this page.
The conversational interface re-poses every one of Norman's questions — what can I do, did it work, what does it know about my model?
Norman's "gulf of execution" — the distance between what you want to do and how to do it — was nearly closed by direct manipulation interfaces. A blank chat box reopens it. There's no signifier for what the model can do. Show the affordances back to the user: examples, suggested prompts, capability hints.
When the model returns an answer, the user's gulf of evaluation reopens — "did it understand me, did it use the right sources?" Norman's feedback principle still applies. Surface the reasoning, the sources, the confidence. A clean answer with no scaffolding feels fast but breaks trust on first error.