Restructuring actions are a broad category that change the structural form of an existing piece of content. This is closely related to Restyling actions that change the surface qualities of content (tone, mood, palette, or voice) while leaving its structure intact, and to Transforming actions which go further, shifting the modality altogether like converting a text summary into a slideshow. Structuring sits in the middle: more structural than restyle, but bounded within the same medium unlike transform.
Examples include:
Condensing: Make the content shorter while keeping the main points intact. Notion AI’s Summarize and Descript’s Remove filler words both let you shrink text or transcripts into something tighter and easier to scan.
Expanding: Add more detail or context to make something fuller. Jasper can turn a short outline into a full blog post, and Notion AI’s Expand command fleshes out bullet points into paragraphs.
Reordering: Change the sequence of ideas or clips without changing their content. Adobe Premiere Pro’s Text-Based Editing lets you reorder video clips by rearranging lines in the transcript, and Copilot Labs can reorganize code functions for clarity.
Perspective shifting: Rewrite for a different audience or point of view. Jasper can convert marketing copy for executives into a customer-facing version, and ChatGPT-style tools can change first-person narration into third-person.
Extraction: Pull out specific elements from a larger body of content. Coda AI and Notion AI both extract action items from meeting notes, while Descript can isolate speaker quotes from transcripts.
Aggregation / merging: Combine multiple sources into a coherent structure. Perplexity AI aggregates results into outlines with references, and tools like Runway let you merge multiple video clips into a highlight reel.
Segmentation: Break a large piece into smaller, structured units. Descript and Riverside both cut long recordings into short clips, and GitHub Copilot can split large functions into smaller ones for maintainability.
Substitution: Swap elements without rewriting the whole thing. Midjourney’s Vary Region swaps part of an image while keeping the rest, and Copilot lets you replace a code snippet with an alternate implementation.
Example from Descript showing the number of use cases for remixing that they support
The value of remixing lies in iteration. Instead of forcing users to re-prompt from scratch, it allows them to build new versions on top of existing drafts. This reduces prompt engineering demands, surfaces more creative possibilities, and helps users treat AI as a collaborator rather than a generator of single shots. Done well, remix flows give people precise levers for reworking content without overwhelming them with raw parameters.
Design considerations
Use presets to make actions clear. Inline actions and related nudges can help expose different options for changing the structure of content and make them easier to apply. Actions like “make shorter” or “remove filler words” are clear and actionable and don't require long, instructional prompts to enact.
Support nuance with sliders. Instead of single-shot actions, allow users to select from pre-set actions to restructure their content, like setting the reading level to re-align to instead of simply offering the action “make more technical.”
Make what changed legible. Show diffs, highlights, or callouts so users can see exactly what was added, removed, or reordered. This helps them trust the system and learn how different restructure actions behave.
Allow user review and selective action. Since changing the structure of content can impact its clarity or continuity, ensure users retain control by verifying changes, selecting from different versions, or at least having an option to undo.
Maintain stylistic tokens. Changing the style or tone of content and the structure itself require separate actions. When changing structure, maintain the overall style and form of the underlying content.
Allow users to target restructured content to specific areas, such as adjusting filler words for individual speakers or converting different sections of an outline into written content to edit and review bit-by-bit.
Restructuring a piece of content can lead to unexpected results. Use variations or branching to create a new version for the user to compare, but allow them to return to the original version.
Examples
FloraFauna shows the paths of connection to remix multiple prompts and inputs on the same canvas
Github Copilot allows restructuring of code via inpainting directly on the canvas. Users are asked to verify the changes before they are adopted
Jasper includes common restructuring tools to adjust the length of selected content.
Jasper supports restructuring directly on the canvas in blend mode, where new attachments and instructions can be added before regeneration at any time.
Microsoft Copilot allows users to restructure content on regeneration with new instructions and prompts for the model to pull from.
Midjourney makes restructuring a core part of its user experience, from adjusting tokens to blending images. Its omni references tool lets users focus generations on specific subjects, and change the subjects or intensity of focus on subsequent actions.
Notion AI includes several inline actions to restructure the selected content or the content on the page
Udio supports restructuring as a core action when working with a pre-composed song. From the canvas, users can use sliders to adjust the intensity of different parameters or the tokens for the regenerated creation
