All patternsWayfinders
Suggestions
Description
Design considerations
Related patterns
Examples

Suggestions

Suggestions are the icebreakers of AI interaction. They help a user learn what they could ask the system to do, and keep the generative conversation moving forward. Generally these appear in a list of three-to-five suggestions that pre-fill the chat input when selected or kick off an action.

AI is most effective once it has information about you. When you start a new thread or open an AI product the first time, the system's knowledge of you is cold. Suggestions provide a way to get some initial context with the user, so the model can start providing more adaptive or personal results.

Forms of suggestions

Depending on the experience, suggestions trend towards one of three forms:

  • ‍Static suggestions show fixed starters for onboarding or first-run experiences. These may vary based on the mode the user has selected, but otherwise are not trained towards the user. Most often these relate to product-specific features, which might prompt you to ask a question about a video (YouTube) or trigger a common action like generating a photo (Canva).
  • ‍Contextual suggestions shift based on the page, document, or content a user is viewing. These update in real time, and can be a form that static suggestions take once a user has interacted with the product. Context might be as simple as a change in mode like offering to make a quiz when switched into study mode (ChatGPT). More complicated examples take context from the canvas or contextual surroundings, like questions specific to a code file (GitHub CoPilot).
  • ‍Adaptive suggestions evolve from individual behavior or team presets, refining over time as the system learns what’s useful. Customized instructions and details might impact how suggestions are generated to be worded (Jasper), or systems might take personal information about the user, their role, and their interests and provide suggestions most relevant to them (GrammarlyGO).

In all cases, even if the suggestion itself is not followed, they encourage users to begin an interaction, allowing the product to better serve their needs.

Design considerations

  • Make suggestions actionable. Clicking on a suggestion should run the prompt, either as a starting point for the user to edit or as an initial prompt that begins the interaction.
  • Leverage contextual clues. Suggestions are most effective when they relate to something the user might reasonably write as a starting prompt. If an attachment is added, draw suggestions from the content. Or, update suggestions when the mode changes to be relevant to common queries, like research questions in research mode or quiz generation in study mode.
  • ‍Scope when suggestions appear. Don’t show them everywhere. Identify moments when users most need direction, such as onboarding, new contexts, or after idle time. In focus-heavy tasks, a blank input may be better. Scope decisions reduce clutter and make the system feel intentional.‍
  • Keep the set small and ranked. Display three to six relevant options, ordered by context or performance. Too many options leads to scanning fatigue. Periodically refresh or hide low-engagement options. Treat the suggestion set as a ranked list that adapts over time.‍
  • Design for first-run learning. Use suggestions to model good prompts during onboarding. Each should demonstrate the system’s capabilities while remaining editable, so users quickly understand how to write effective requests.‍
  • Respect safety and cost. If a suggestion could trigger data access, publishing, or heavy compute, show a lightweight preview or confirmation. Users should understand implications before committing.

Related Patterns

Inline Action

Suggestions are helpful if the user wishes to continue in a linear flow of iteration, but inline actions can proactively suggest divergent exploration and new branches for discovery.

Follow up

When targeting suggestions based on the user's previous prompt or the context of the interaction thus far, position suggestions as follow ups. This encourages the user to share their intent, needs, and references up front for the AI to work with.

Examples

‍

Canva’s AI-generated search suggestions map keywords based on the user’s search query.
Gemini search can read the information on the page and propose suggested questions the user might have without them having to dig.
Granola provides standard suggested recipes that reflect common questions someone might have while reviewing a transcribed meeting.
LinkedIn suggestions draw from the context of the post they sit beneath, increasing the likelihood that they are relevant to the viewer.
Miro calls context from the content of the canvas when suggesting actions to the user.
Typeform’s AI suggestions are provided before the user has given any context about their intended form, so they are likely to be irrelevant to the user.

© Emily Campbell 2025 | Sharable under CC-BY-NC-SA | Contact

Sharable under the Creative Commons CC-BY-NC-SA license