wunder beta📘 Prompt UX: AI Product Management
A bare text box gives the user infinite freedom and zero guidance, which is paralyzing rather than liberating.
What you’ll learn
- Lesson 1: The Blank-Canvas ProblemExplain the blank-canvas problem and choose entry-experience patterns (examples, templates, suggested prompts, starter chips, progressive disclosure) that reduce it.An empty prompt box offers infinite freedom and zero guidance, which paralyzes users and drives abandonment; this is the blank-canvas problem. The fix is to convert the open canvas into a guided start using patterns that teach and seed the first action. Examples demonstrate phrasing and scope by imitation; templates give fill-in-the-blank scaffolds; suggested prompts and starter chips offer one-tap entry, especially on mobile. Treat the prompt field as a primary interface whose placeholders and scaffolding shape behavior, and use progressive disclosure to reveal advanced controls only after a first result. Match the amount of guidance to user expertise, since novices and experts need different levels of scaffolding.
- Lesson 2: Communicating Capability, Uncertainty, and TrustDesign interfaces that set accurate expectations, communicate AI uncertainty honestly, and calibrate user trust to actual per-task reliability.Users form a mental model of an AI feature from its first screens, so explicitly signaling what it is good at and not designed for steers requests toward well-supported tasks. Because generative models are probabilistic and can be confidently wrong, the interface should convey uncertainty honestly through hedged language, verifiable citations, and clear AI-generated labels rather than false confidence. The goal is calibrated trust, where reliance matches actual reliability and is earned per task, avoiding both over-trust that accepts errors and under-trust that abandons a useful tool. Explanations should scale with decision stakes, and disclosing AI involvement plus mitigating harm in high-stakes flows are first-class design responsibilities.
- Lesson 3: Latency, Errors, and Human-in-the-LoopApply patterns for perceived latency (streaming), graceful error/empty/refusal states, editability, human-in-the-loop control, and efficient iteration.Generative responses are slow and fallible, so the response experience must manage perception and failure. Streaming output reduces perceived latency by letting users read immediately rather than wait behind a spinner. Graceful error states explain what happened, preserve input, and offer Retry, Rephrase, or Adjust scope; empty states teach with examples; refusal states explain the limit and suggest an acceptable alternative. Making output editable keeps a human in control and lets users fix imperfect results instead of abandoning the tool, serving both safety and usability. Because first prompts rarely succeed, an efficient refinement loop, quick actions, preserved context, alternatives, and version history, captures most of the feature's value.
- Lesson 4: Guided Project: Redesign an AI Feature (Mastery)Synthesize all course principles into a UX redesign of a high-abandonment AI writing assistant, justifying each decision and tying it to a measurable outcome.This capstone applies the full toolkit to a writing assistant where most users open and close it without typing and the few who try rarely return. Step one attacks the blank canvas with an entry experience of examples, templates, and starter chips, matched to user expertise and measured by the share reaching a first useful output. Step two designs the response experience: streaming to cut perceived latency, AI-generated labeling and verifiable cues to calibrate trust, graceful error/empty/refusal states that preserve input, editable output with accept/edit/discard for human-in-the-loop control, and a cheap refinement loop. Every decision is justified and tied to a metric such as return rate, edit rate, or task completion.
Questions this course answers
Users open your new AI feature, see only an empty text box, and most close it without typing anything. Which framing best explains this and points to the fix?
The empty box gives infinite freedom and zero guidance, the blank-canvas problem; the remedy is in-context guidance like examples and templates. There is no evidence users reject AI outright or that latency is the issue (they never typed), and dumping configuration on users before any payoff worsens abandonment.
Your team debates whether to teach prompt-writing with a written rules document or a small gallery of worked example prompts. Which is generally more effective for getting users started, and why?
Concrete examples teach phrasing and scope faster than abstract rules because people imitate and adapt them, and they implicitly signal boundaries. Long rule documents are rarely read and do not demonstrate phrasing; leaving users to self-discover is the blank-canvas trap; localization cost is not the deciding usability factor here.
A mobile AI feature needs to lower the cost of the very first input for novice users. Which pattern is the best fit?
Starter chips are ideal on mobile because one tap launches a useful first turn with no typing, directly lowering articulation cost for novices. An empty box is the problem itself; a mandatory settings panel or required tutorial adds friction before any payoff and increases abandonment.
Why should a product manager treat the prompt input field as a primary interface deserving full design rigor?
The prompt is how users express intent, so its design directly drives whether they succeed, which is why guidelines treat the entry point as core. The other options are false: there is no universal legal mandate here, box size alone does not drive engagement, and the field is far from cosmetic.
Before users try your AI feature, what is the main reason to clearly state what it is good at and not designed for?
Early, explicit capability and limit signaling builds an accurate mental model, steering users to well-supported tasks and calibrating trust. It is interaction design, not marketing filler; it is not meant to suppress feature use; and disclosure cannot prevent the model from making mistakes.
Your AI sometimes produces fluent but factually wrong answers, and users accept them without checking. Which design response best addresses this while maintaining honesty?
Honest uncertainty cues plus verifiable sources and labeling let users calibrate trust and catch errors. A fabricated confidence score is dishonest and dangerous; hiding uncertainty encourages over-trust; removing citations strips away the very means users have to verify.
Grounded in trusted sources
- Google PAIR, People + AI Guidebook, https://pair.withgoogle.com/guidebook/
- Nielsen Norman Group, AI UX, https://www.nngroup.com/topic/artificial-intelligence/
- OpenAI, Prompt Engineering Guide, https://platform.openai.com/docs/guides/prompt-engineering
- Microsoft, Guidelines for Human-AI Interaction, https://www.microsoft.com/en-us/research/publication/guidelines-for-human-ai-interaction/
Every Wunder lesson is built from real, reputable sources — never invented.
Related courses
Wunder is a personalized learn-anything platform — tell it any topic and it builds a beautiful, fact-checked course in minutes, with narration, a knowledge check, and a college-style University track.
© 2026 Wunder Learning LLC · Terms & Privacy