5.2.2 Strategies for Individual Editing Goals.
Our analysis also sheds some light on how participants achieved individual composition and editing goals, including adding, reviewing, reorganizing, rephrasing, and removing content.
Adding Content. All 11 participants used dictation to compose text, through creating a new Ramble, as their main way of adding new content. Seven participants (P1, P3, P4, P5, P7, P8, 12) also utilized Respeaking to add content to an existing Ramble. P1 used Manual Split to separate a specific sentence out of a Ramble so that they could add onto it.
Reviewing content. Eight participants used either Semantic Zoom, Highlighting Keywords, or Regenerating Summaries to explicitly help review their composition. Four participants (P4, P8, P10, P12) used Semantic Zoom to help review during their process, while three (P2, P7, P11) used Semantic Zoom to only review at the very end of their composition task. One participant (P12) also used Highlighting Keywords by itself as a form of review, while P2 and P10 used Highlighting Keywords combined with Regenerating Summaries to review.
Reorganizing Content. As expected, participants made use of the Manual (onscreen keyboard-based) and Semantic (LLM-based) Split and Merge functions to segment content. But we also observed interesting mix-and-match strategies combining manual and semantic operations. Three participants (P1, P4, P12) used Semantic Split first, before using Manual Merge on some of the resulting Rambles to achieve a partially Semantic, partially Manual Split. P3 used Semantic Merge and then Manual Split to control some of the uncertainty of what they would get from Semantic Merge. P1 also paired Manual Split and Manual Merge to more precisely resegment their Ramble (by moving the last sentence of the first Ramble into the second Ramble). More uniquely, P10 got rid of all their segmentation through a combination of Manual Merge and Semantic Merge, and P4 experimented on a Ramble through Manual Merge, then Semantic Split, and then used Semantic Merge on the resulting pieces.
Furthermore, three participants (P1, P3, P4) utilized Magic Custom Prompt to attempt to split their composition: P3 (“make into bulleted list”) and P4 (“Reformat like a design proposal”) did so at the beginning of the task, while P1 (“Can you… organize it into paragraphs”) did so at the end of their task. As Magic Custom Prompt is designed to only modify a single Ramble and Rambles do not store paragraph breaks, these prompts would just create a block of text within a Ramble, and participants would go on to either replace through Respeak Ramble (in the case of P3) or segment it through Semantic Split.
Seven out of 11 participants (P1, P2, P3, P4, P9, P12) rearranged their Rambles through drag and drop Reordering, with P1 and P4 also altering the order of text by using Manual Merge on non-consecutive Rambles. Two participants (P1, P3) also more specifically rearranged content within their Ramble: P1 did so through Manual Split and then either reordered or did a non-consecutive Manual Merge, while P3 asked Magic Custom Prompt (on a single Ramble) to “reflow this text to make sense”.
Rephrasing Content. All participants used keyboard editing to make minor changes like fixing typos or rewording phrases. Ten out of 11 participants utilized Magic Custom Prompt to do a more broad rephrasing. Three participants (P2 P8, P12) only sought specific style changes like “make it more formal” or “like a tumblr post”. Five participants (P1, P5, P7, P9, P11) looked for a more general revision of the text through less specific requests like “revise text,” “clean the text,” or “make this writing more readable”, while two participants (P3, P4) did both. Furthermore, five participants tried to achieve a more uniform style across their entire composition: three participants (P2, P3, P8) used similar prompts across all their Rambles, while two participants (P1, P5) used Manual Merge to combine all their Rambles into a single Ramble so that they could apply Magic Custom Prompt onto their entire text in one go. Two participants (P3, P12) also used Respeaking to replace entire Rambles.
Removing Content. Ten out of 11 participants deleted content from their composition. Seven participants (P1, P2, P3, P4, P5, P8, P12) utilized the Delete Ramble functionality, and three of those participants (P3, P5, P8) specifically combined Manual Split with Delete Ramble to delete words and sentences from the beginning or end of a Ramble. Three participants (P1, P2, P11) utilized keyboard editing to delete whole sentences from their Rambles.