Onboarding, Empty States, and Suggested Prompts
Solve the blank-box problem: teach capabilities in context, suggest starting points, and help users form an accurate model of what the AI can do.
An empty text box with "Ask me anything" above it is one of the hardest interfaces you can hand a new user. It tells them the feature can do something, but not what, not how well, and not how to phrase a request so it works. Most people respond in one of two ways: they type something vague and get a vague answer, or they test it with something it can't do, it fails, and they decide the feature isn't for them. Either way, the first few interactions set a mental model that's hard to change later.
This is the blank-box problem, and it isn't solved by a welcome tour. It's solved by teaching capabilities at the moment they're useful, suggesting starting points that fit what the user is doing right now, and handling the first failure so it teaches rather than discourages. In this lesson you'll design all three for Deskmate's assistant panel, and learn how to measure whether onboarding is working.
Why the first five interactions matter
Users build a working theory of what an AI feature can do from the first handful of things they try. With ordinary software, the interface itself teaches capabilities: there's a button for each thing the product does. With a text box, the user has to discover the capabilities by experiment, and every experiment updates their theory.
That theory goes wrong in predictable ways:
- Underestimation. A user tries a simple request, gets an adequate answer, and never discovers the feature can do much more. They use it as a slightly better search box.
- Overestimation. A user assumes it knows everything the product knows, asks about something it can't see (a billing record, an internal note), and gets a confident wrong answer or a confusing refusal.
- Wrong phrasing. A user writes requests the way they'd write a search query ("refund policy annual") and gets mediocre results, and concludes the feature is mediocre.
Good onboarding isn't about showing everything up front. It's about making the first interactions likely to succeed, and making the boundaries visible early, so the user's theory matches reality.