Mental Models & the First Five Minutes

A chatbot, an oracle, an agent, and an autocomplete are four different machines, and users hold four different sets of expectations for them. A chatbot is a conversation partner — you expect turn-taking, memory of what you just said, and tolerance for vagueness. An oracle is a question-answering authority — you expect a single correct answer and you punish it for being wrong the way you’d punish an encyclopedia. An agent is a delegate — you expect it to act, and you expect to be asked before it does anything expensive. Autocomplete is a reflex — you expect speed, low ceremony, and the right to ignore it without cost.

The trouble is that people import the wrong model constantly. They treat ChatGPT like an oracle and get burned by a confident fabrication. They treat GitHub Copilot’s ghost text like an authority instead of a draft and ship its bugs. They treat an agent like a chatbot, chatting past the moment it started executing. None of these are user errors — they are borrowed models the product failed to replace.

Your first design decision is therefore diagnostic: which machine are you actually shipping, and which machine will people assume? The gap between those two answers is your onboarding scope. Everything else in this chapter — contracts, starters, form, first runs — is machinery for closing that gap before the user pays for it.

The fastest way to replace a borrowed model is to state your own. Microsoft’s HAX guidelines put it as rules one and two for a reason: make clear what the system can do, and make clear how well it does it. Most teams ship only the flattering half. “Summarizes your meetings” is marketing; “built from your meeting audio plus the notes you type — it won’t capture what nobody said out loud” is a contract, and Granola’s framing along those lines is why its users show up as participants rather than bystanders waiting on magic.

The contract belongs where the work happens, not in a help-center article. ChatGPT states its knowledge cutoff at the moment you ask about recent events. Claude tells you when it can’t open a link, right in the reply. A limit surfaced when it becomes relevant is remembered; a capabilities page read once during signup is not. And self-identification is part of the deal — people calibrate differently the instant they know they’re talking to a machine, and every product that has blurred that line has spent more repairing the trust than the ambiguity ever earned.

The test: when a new user hits their first three failures, do they say “right, it told me it couldn’t do that,” or “I thought this thing was supposed to be smart”? The first is a contract working. The second is a lawsuit you filed against yourself.

An empty text box is a test, and most people fail it. Faced with a blinking cursor and infinite possibility, a new user asks something trivial — “tell me a joke,” “what can you do” — gets a trivial answer, and concludes the product is trivial. That’s the blank-box tax: you outsourced your value proposition to the imagination of someone who hasn’t decided to trust you yet, and they billed you for it in churn.

The tax is avoidable because the first prompt is yours to design. ChatGPT and Claude greet new sessions with concrete example prompts, and the ones that work are specific enough to try verbatim — “summarize this PDF” converts; “get creative!” confirms the suspicion that the product doesn’t know what it’s for. Perplexity seeds its home screen with suggested and trending questions, which does double duty: it starts a first query and teaches, by example, that this is a question-answering product rather than a chat toy.

The stronger move is to skip the box entirely and go to where the problem already lives. Notion AI appears when you hit space in an empty block; Gmail offers to draft a reply on the message you’re already reading; Figma surfaces AI actions on your current selection. Contextual entry points collapse the hardest step — translating your situation into a query — because the entry point carries the context for you. The person’s first result is relevant by construction, and relevance on the first try is worth more than any tour.

Before a user reads a word of your copy, the shape of the interface has already told them what kind of machine this is. Ghost text — Copilot’s gray inline suggestion — says “draft, optional, ignorable”; nobody writes a prompt to ghost text, and nobody blames it for being imperfect, because its form pre-negotiated its authority. A chat bubble says “conversation partner”: it invites open-ended requests, follow-ups, and the assumption of memory. An inline action on a selection — “translate this,” “make shorter” — says “tool”: bounded, predictable, done in one shot. Same model behind all three, radically different contracts.

This means form is your cheapest and most honest expectation-setting instrument, and misusing it is expensive. Put a chat bubble on a system that can’t handle follow-ups and you’ve promised a conversation you can’t deliver. Render probabilistic guesses in the same authoritative layout as retrieved facts — as the first wave of AI search summaries did — and you’ve borrowed the credibility of the encyclopedia form for outputs that couldn’t cash it.

Modes extend the same logic to depth. When ChatGPT separates fast models from slow reasoning modes, or Perplexity splits quick search from deep research, the mode picker itself teaches the user that quality is a dial, latency is a choice, and the system’s effort is legible. Suggested actions teach scope the same way: the verbs you offer on a selection are a menu of what this machine is for. Design the form first — the copy is just captioning it.

The goal of an AI first run is one real success on the user’s own material, not a tour of forty features. A guided first run should route people toward tasks the system is genuinely good at — even if that means hiding the impressive-but-flaky capabilities for now — because the first result calibrates everything after it. One excellent summary of the user’s actual document beats ten hypothetical capabilities on a carousel, every time.

The highest-leverage first-run moment in most AI products is context attachment. The instant someone drops in their own PDF, connects their codebase, or points the system at their real meeting, two things happen at once: the output quality jumps, and the mental model snaps into focus — “it works on my stuff.” Claude’s file attachment, Cursor indexing your repository, NotebookLM built entirely around your uploaded sources: each makes attaching context the first move, not an advanced feature, because generic output from a context-free prompt is the weakest possible first impression.

Then get out of the way. Scaffolding that lingers becomes an insult — showing generic starter prompts to someone with fifty sessions of history reads as the product forgetting who they are. Good onboarding is measured by how gracefully it disappears: starters yield to the user’s own patterns, coach marks never repeat, and the guided path quietly becomes the default path. If your onboarding is still visible in week three, it isn’t onboarding — it’s clutter with a welcome message.

Every claim made about your AI anywhere — the keynote, the landing page, the app-store screenshots, the CEO’s post — is a deposit into an expectation ledger that your interface has to pay out, interaction by interaction. Users don’t distinguish between what marketing promised and what design shipped; they experience one product and one gap. When the demo shows a flawless agent booking travel end-to-end and the shipped feature needs three corrections to draft an email, the interface eats the difference, and it eats it as distrust.

The field has run this experiment repeatedly. Overpromising launches — humanoid-assistant hype, “it just does it for you” demo reels — generate first sessions where the user’s opening move is beyond the system’s actual envelope, which guarantees the first result is a failure against an inflated baseline. Compare the products that underclaim and overdeliver: Copilot marketed itself as a pair programmer offering suggestions, and users who found it writing whole functions felt they’d discovered a bonus rather than caught a lie.

The practical discipline is to treat expectation-setting as a designed surface with an owner. Audit every public claim against what the median first session actually delivers, and file the gap as UX debt — because that’s what it is: debt the interface services with apology patterns, disclaimers, and churn. Designers should sit in marketing reviews for the same reason they sit in eng reviews. The mental model gets built wherever the user first hears about you; you might as well be in the room.