All patternsPrompt actions
Summary
Description
Design considerations
Related patterns
Examples

Summary

The summary action condenses and organizes information to make it easier to understand, recall, and act on. This action may span one or many sources, but it remains faithful to the original material with an emphasis on isolating key points or relevant information.

When data is presented as a summary, users should have confidence that the model is accurately capturing and compressing the source material, and not interpreting it or injecting opinion. This makes the resulting summary more reliable. The synthesis action can be helpful, but users need to be able to confidently distinguish between automated summaries of human-created content, and machine-generated interpolations.

To ensure users can trust that the summary is an objective reflection of the source material, combine summaries with governors that help users retain agency in the experience.

  • Citations help users understand which sources map to which parts of the summary
  • References guide users to the source material where they can review and verify the cited information
  • Follow ups recommend additional ways to interact with the source material, such as providing and alternative view or researching terms and ideas.

Design considerations

  • ‍Prioritize fidelity over brevity. The goal of a summary is not compression at all costs, but clarity without distortion. Shorter must still mean true. Ensure summarization captures all key claims and context so the user can trust that nothing essential has been lost.‍
  • Expose scope and source. Ensure the viewer can identify what is being summarized (a document, chat thread, page section, or selection). Make the boundary explicit so users know what’s included and what’s not. Cite or link to the original material so details can be verified.‍
  • Provide automatic summarization with caution. Automatic summaries can speed scanning for long or routine material like email threads, meeting notes, or chat histories, but they’re risky for dynamic or opinionated content such as news articles, search results, or social feeds. In those settings, auto-summaries may distort tone or prematurely frame interpretation. Always make it clear whether a summary is AI-generated, and give users the option to expand, refine, or disable it entirely to encourage critical discovery.‍
  • Match granularity to context. A chat recap, a research digest, and a paragraph summary each demand different levels of compression. Offer presets like “short,” “detailed,” or “key points.” Users should be able to adjust summary depth rather than accept a single fixed form.‍
  • Make attribution automatic. When summarizing sourced material, attach citations as inline references or backlinks. Citation links reinforce credibility, allowing users to verify critical information and guarding against hallucination.
  • ‍Preserve the original language when precision matters. In legal, scientific, or contractual contexts, verbatim phrasing can be essential. Provide a “quote and condense” mode that lists key passages with light framing instead of rewriting.‍
  • Audit and test for omission bias. Periodically compare machine summaries against expert-written ones to detect consistent blind spots—such as omitting counterpoints or minority perspectives. Adjust weighting or training data accordingly.

Related Patterns

References

Listing the references that compose the source material for the summary allows users to verify the summarized content and identity what content is most relevant to their query or needs.

Citations

While summaries provide an overall summary of a set of information, exposing the most pertinent points, citations provide paths back into the source of those points to help users drill into the source material more effectively and verify what was extracated.

Synthesis

Use the synthesis action when the goal is to integrate or interpret across multiple sources, not just condense them.

Examples

‍

Apple’s attempt to summarize conversations can lead to surprising results
Bloomberg shows how AI Summaries can be integrated into products that otherwise are not oriented towards AI as a core feature. Consider deep linking from summaries into the article to encourage deeper engagement and thoughtful reading
Figjam plays with the Canvas surface, using summary to interact with other elements, like a set of sticky notes
Hubspot relieves cognitive load by allowing users to summarize a large set of information 
Notion uses summary in two different ways: inline, to remix text into a summarization, and applied to the full page, to generate a summary of all content
Perplexity’s success largely stemmed from their re-invention of the search paradigm, focusing on summarizing results instead of providing a ranked list
Auto-generated summary cards have being ubiquitous. To make them more effective, include citations as “proof” to lead users to understand how the summary was derived
Salesforce helps support agents by given them a nudge to summarize the conversation thus far, benefitting the agent as well as the user, who may not need to repeat information
Slack has added the ability to summarize a thread of content. Adding deep linking to individual posts via citations would make it even easier to follow up on what needs action vs. what is noise
Zoom uses the summary pattern in live meetings to let users quickly catch up on context they may have missed

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

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