Prompt suggestions

Prompt suggestions, sometimes called starter prompts, are clickable examples that sit near an AI input box: "Summarize this thread," "Draft a reply that declines politely," "What changed since last week?" NN/g defines them as "system-generated hints that guide users in forming queries or commands for AI tools." They show up before the first message, as autocomplete while typing, and as follow-up questions after an answer.

Why it matters

An empty text box gives no clue what the system can do or how to ask. Writing a good prompt is a skill, and many people do not know where to start. Suggestions turn that recall task into recognition: people pick something close to what they want and adjust it.

NN/g's Prompt Suggestions article (Moran and Neusesser, 2025) describes three types: use-case suggestions that show what is possible, prompt autocomplete that speeds up typing, and followup questions tied to the current conversation. Apple's generative AI guidelines recommend the same idea: "offer diverse, predefined example inputs that hint at what's possible for a feature."

How to apply it

Do:

• Make suggestions specific to the context. NN/g found that "broad or generic prompt suggestions are rarely effective." In a meeting-notes tool, "List action items with owners" beats "Ask me anything."
• Tie them to the page or object the user is looking at. NN/g's chatbot guidelines advise tailoring the opening message and suggestions to the current page, and offering suggested questions as buttons, not plain text.
• Update follow-up suggestions as the conversation moves, and drop ones the user has already ignored.
• Decide whether a suggestion sends immediately or fills the input for editing. Cloudscape's support prompts pattern uses editable prompts for getting started and one-tap prompts for quick next steps like "Show more details."
• Use suggestions to teach good prompting, for example by including the kind of detail (audience, length, tone) that improves results.

Don't:

• Suggest things the system cannot do well. Each suggestion is a promise about capability, and a failed example damages trust quickly.
• Show a long list. A few varied, relevant options are easier to scan than a menu of twenty.
• Write suggestions in the AI's voice ("I can help you...") when they will be sent as the user's message.

Common mistakes

• Static, generic examples. The same three suggestions on every screen, regardless of context, are quickly ignored.
• Suggestions as decoration. Clicking one produces a weak answer because nobody tested the prompt.
• Hiding them after the first message. Follow-up suggestions are often more useful than starters, because they show the next step from a real answer.
• No way to edit. One-tap prompts that cannot be adjusted force people to retype a near-match from scratch.
• Dead ends. A suggestion that leads to "I can't help with that" teaches people the feature is unreliable.