Mental Model

A mental model is the user's internal map of how a system operates. For AI tools this map must form in under sixty seconds because the interface hides the machinery behind a single prompt box. The map answers four questions the user brings with them. What exactly can this thing do and what is it bad at. How am I supposed to talk to it. What does a good output look like. And what should I type first. The teams that solve AI onboarding treat those four answers as the entire design problem for the first run. They do not treat onboarding as a checklist of feature announcements.

It is not a feature tour. It is not a carousel of screenshots. It is not a tooltip that explains the model architecture. Those are all attempts to push information at the user. A mental model is what the user actually believes after they close the browser tab. If the belief is inaccurate the user will generate weak prompts, misjudge outputs, get frustrated, and leave. The entire point of great AI first run design is to install an accurate model before the user has time to build a wrong one.

ChatGPT showed how to do this on day one in late 2022. The landing page used a three column layout. Examples on the left. Capabilities in the center. Limitations on the right. The page answered every major question before the user typed a single character. Users walked away knowing the tool was powerful but not magic, that it could remember context but might hallucinate, and that they should start with a clear request. The model installed in one screen. Most copycat products that followed still have not matched that clarity.

Cursor took the opposite approach in 2024. The first run is not a page about Cursor at all. It is a request to open one of your own codebases. Once the folder loads the AI surfaces through familiar editor shortcuts. The mental model forms around the user's real files instead of a fake demo project. This eliminates the transfer problem that kills every sandbox based onboarding. The user never has to wonder whether the tool will work the same way on their actual work. They see it on the first try.

Claude.ai used example prompts as the primary teacher. The homepage in 2023 was a collection of well written prompts that demonstrated tone, length, and structure. Clicking one launched a real conversation with streaming output. The example carried the entire mental model. The user saw what good input looked like, what good output looked like, and what the AI was willing to do. No separate tutorial required.

Perplexity solved the success state problem by showing a complete answer to a trending question the moment the page loaded in 2022. Users saw citations, related questions, and the full layout of a finished research session. They did not have to imagine what done looked like. The design removed the paralysis that comes with a blank prompt bar and no reference. Granola tied the first run to a real world event. Their onboarding flow was one line. Connect your calendar. The product then waited until the user joined their next real meeting. The first AI output happened on actual notes from an actual call. The mental model formed in the context that mattered instead of a sterile demo. Linear proved you do not need a standalone AI onboarding at all. They embedded the AI features inside the flows users already followed. When drafting an issue the AI offered to polish the text. When viewing a backlog it suggested priorities. The mental model grew organically without context switching or special modes. The user learned the tool while doing their actual job.

These examples all obey the same three rules. Show the capability surface instead of describing features. Reach first prompt in under thirty seconds. Make the success state visible in under two minutes. Break any of those rules and activation collapses no matter how slick the visual design. Adobe Firefly followed the same logic in 2023 by rendering multiple generation styles on the landing page so users immediately understood quality variance and prompt influence before they typed their first description.

The patterns that destroy mental models are just as clear. Tutorial walls that force five clicks before the prompt appears teach users nothing except how to click next. Modal carousels full of feature promises set expectations the actual product cannot meet. Form gates that demand company size and role before any value appears feel like a tax. Each pattern bets that explanation can replace demonstration. Every shipped product that uses them proves the bet wrong.

Apply this mental model lens when you are designing the first run for any prompting based AI tool. The approach is required whenever the user must steer a non deterministic system toward useful output. Skip it for AI that stays invisible such as automatic transcription cleanup or background ranking improvements. Those earn their place through repeated quiet value not through explicit teaching. Also skip it for products whose users are highly technical and motivated. They will build their own models through experimentation faster than any designer can script the experience.

The job is to make the invisible shape of the AI visible immediately. Show the bounds. Teach the language. Display the win condition. Trigger the first real attempt before the user loses interest. Ship an honest mental model on the first screen or watch users build a fantasy version that your product can never satisfy.