cs.HCApr 8, 2026
SaveNarrix: Remixing Narrative Strategies from Examples for Story Writing
Organizations: Cornell University Ithaca, NY, USA · Adobe Research San Jose, CA, USA
Abstract
Experienced storytellers decompose stories into local narrative strategies and how these strategies shape higher-level arcs. This decomposition helps writers recognize patterns in others' work and adapt those patterns to tell new stories. Novices, however, struggle to identify these strategies or to reuse them effectively. We present Narrix, a novel writing tool that helps novice writers recognize narrative strategies in example stories and repurpose these strategies in their own writing. Narrix analyzes strategies in example stories, highlights them with color-coded lexical cues and explanations, and situates them on an interactive story arc for exploration by emotional shifts and turning points. Writers then drag strategies onto multi-dimensional tracks and apply block-scoped edits to revise or continue their drafts through controlled generation steered by specified strategies. Through a within-subjects study (N=12), Narrix showed improved participants' retention, confidence, and creative adaptation of narrative strategies compared to a baseline chat-based writing interface.
Figures & tables
Figure 1. Three main features of Narrix : (A) Surfacing Narrative Strategies: Narrix automatically extracts narrative strategies from example stories, highlighting relevant lexical cues in color and explaining their function in the story. (B) Exploring Narrative Strategies: Users can retrieve example story content and their strategies based on evolving higher-level storytelling intents (e.g., emotional shifts, turning points) via an interactive story arc. (C) Remixing Narrative Strategies: Users can drag and drop selected strategies onto multi-dimensional tracks (e.g., character, plot, linguistic) to steer the AI in revising or continuing their story by integrating those strategies. Three main features of Narrix: (A) Surfacing Narrative Strategies: Narrix automatically extracts narrative strategies from example stories, highlighting relevant lexical cues in color and explaining their function in the story. (B) Exploring Narrative Strategies: Users can retrieve example story content and their strategies based on evolving higher-level storytelling intents (e.g., emotional shifts, turning points) via an interactive story arc. (C) Remixing Narrative Strategies: Users can drag and drop selected strategies onto multi-dimensional tracks (e.g., character, plot, linguistic) to steer the AI in revising or continuing their story by integrating those strategies.
Figure 2. Narrix ’s user interface , organized into three views: (A) Markdown Editor for drafting the user’s story, with a mode switch (D) to toggle to the Story-Arc Inspector (see Fig. 3 ); (B) Browser for exploring example content blocks and their narrative strategies; and (C) Remix Tracks for layering strategies across creative dimensions. In the Browser, users can choose different card views (E), open strategy explanation panels (F), and perform semantic or word searches for strategies (G). Users can drag strategies from example cards onto relevant creative-dimension tracks (I) below the story track (H). Strategy tiles in the Remix Tracks can be resized to span multiple blocks (J). Users can steer AI to revise blocks (K) or continue the story (L) using the selected strategies. Narrix's user interface, organized into three views: (A) Markdown Editor for drafting the user’s story, with a mode switch (D) to toggle to the Story-Arc Inspector; (B) Browser for exploring example content blocks and their narrative strategies; and (C) Remix Tracks for layering strategies across creative dimensions. In the Browser, users can choose different card views (E), open strategy explanation panels (F), and perform semantic or word searches for strategies (G). Users can drag strategies from example cards onto relevant creative-dimension tracks (I) below the story track (H). Strategy tiles in the Remix Tracks can be resized to span multiple blocks (J). Users can steer AI to revise blocks (K) or continue the story (L) using the selected strategies.
Figure 3. Narrix ’s Story-Arc Inspector , which juxtaposes the user’s evolving story arc (red line) with arcs from example stories. Each point represents a story block, positioned by story progression (x-axis) and the protagonist’s emotional valence (y-axis). Hovering over a point (B) reveals its details and highlights the corresponding card in the Browser; brushing the points (C) selects a range of story progression and emotion. The system automatically overlays the most similar example arc in yellow (D) and optionally additional arcs in green (E) toggled by user click for comparison. Users can filter story blocks by turning-point type (F) to focus on strategies serving specific narrative purposes. Narrix's Story-Arc Inspector, which juxtaposes the user's evolving story arc (red line) with arcs from example stories. Each point represents a story block, positioned by story progression (x-axis) and the protagonist's emotional valence (y-axis). Hovering over a point (B) reveals its details and highlights the corresponding card in the Browser; brushing the points (C) selects a range of story progression and emotion. The system automatically overlays the most similar example arc in yellow (D) and optionally additional arcs in green (E) toggled by user click for comparison. Users can filter story blocks by turning-point type (F) to focus on strategies serving specific narrative purposes.
Figure 4. Narrix ’s Track Selection panel for adding tracks of different creative dimensions to the Remix workspace. Users can choose from eight predefined dimensions, e.g., Plot, Character, Information, Emotional, Linguistic, Pacing, Thematic, and Engagement. Narrix's Track Selection panel for adding tracks of different creative dimensions to the Remix workspace. Users can choose from eight predefined dimensions, e.g., Plot, Character, Information, Emotional, Linguistic, Pacing, Thematic, and Engagement.
Figure 5. Narrix ’s Reflective Comparison view, which contrasts: (A) the original example content and its highlighted lexical cues for the selected strategy, (B) the revised user content with corresponding cues, and (C) a generated comparison explaining how the strategy is realized in both texts and what differences or similarities exist. Narrix's Reflective Comparison view, which contrasts: (A) the original example content and its highlighted lexical cues for the selected strategy, (B) the revised user content with corresponding cues, and (C) a generated comparison explaining how the strategy is realized in both texts and what differences or similarities exist.
Figure 6. Narrix ’s technical pipeline for processing narrative examples. (A) Segment each example story into content blocks. (B) Infer narrative strategies for each block, including a concise label, explanatory description, and highlighted lexical features. (C) Classify each block for the presence of turning points using fine-tuned models. (D) Identify the story’s protagonist. (E) Infer the protagonist’s emotions per block, map them to valence scores using the NRC VAD Lexicon, and sketch a story arc. (F) Compare the user’s arc to example arcs using DTW to retrieve the most similar narrative trajectory. Narrix's technical pipeline for processing narrative examples. (A) Segment each example story into content blocks. (B) Infer narrative strategies for each block, including a concise label, explanatory description, and highlighted lexical features. (C) Classify each block for the presence of turning points using fine-tuned models. (D) Identify the story's protagonist. (E) Infer the protagonist's emotions per block, map them to valence scores using the NRC VAD Lexicon, and sketch a story arc. (F) Compare the user's arc to example arcs using DTW to retrieve the most similar narrative trajectory.
Figure 7. The distribution of events related to exploring, learning, and remixing narrative strategies across the normalized time span. Exploring Narrative Strategies (yellow circles) includes actions such as interacting with the interactive story arc and applying filters. Learning Narrative Strategies (blue circles) includes actions such as checking and reading stories, strategies, explanations, and lexical hints. Remixing Narrative Strategies (purple circles) includes actions such as manipulating strategies and tracks within the Remixer, as well as revising or continuing the story. The distribution of events related to exploring, learning, and remixing narrative strategies across the normalized time span. Exploring Narrative Strategies (yellow circles) includes actions such as interacting with the interactive story arc and applying filters. Learning Narrative Strategies (blue circles) includes actions such as checking and reading stories, strategies, explanations, and lexical hints. Remixing Narrative Strategies (purple circles) includes actions such as manipulating strategies and tracks within the Remixer, as well as revising or continuing the story.
| Scales | Narrix | Baseline | Statistics | ||||
|---|---|---|---|---|---|---|---|
| M | SD | M | SD | W | p | ||
| NASA Task Load | Mental | 4.42 | 1.44 | 4.00 | 2.05 | 23.00 | .525 |
| Physical | 4.00 | 1.86 | 3.00 | 1.28 | 34.50 | .170 | |
| Temporal | 3.83 | 1.80 | 4.00 | 1.86 | 25.50 | .878 | |
| Performance | 5.42 | 1.08 | 4.17 | 1.12 | 56.00 | .043* | |
| Effort | 3.67 | 1.23 | 4.33 | 1.16 | 2.00 | .089 | |
Table 1. Survey results of perceived experience on task workload, creativity support, and AI system experience under two conditions, where the Wilcoxon signed-rank paired t-test W-values and p-values (*: , **: , ***: ) are reported.
Figure 8. Participants’ responses to a 7-point Likert-scale questionnaire, assessing perceived support across five design goals (see § 3.1 ) in both the Baseline condition and our system. A horizontal stacked bar chart comparing user survey responses for Narrix (our system) versus a baseline across five design goals. The goals are: DG-M, seeing narrative strategies in examples; DG-C, understanding strategies in context; DG-S, applying strategies in a story; DG-R, reflecting on use of strategies; and DG-E, experimenting with strategies. For each goal, responses are shown on a 7-point Likert scale from 1 (strongly disagree) to 7 (strongly agree). Bars for Narrix are dominated by higher agreement values (5–7), while baseline responses cluster more toward disagreement and neutral. Overall, the figure indicates that Narrix is perceived as more effective across all five design goals.
Figure 9. Bar plots illustrating the statistical metrics of participant performance in (A) learning narrative strategies and (B) remixing narrative strategies, where the t-values from the Student’s paired t-test, W-values from the Wilcoxon signed-rank paired test, and p-values (*: p <.05, **: p <.01, ***: p <.001) are reported. Error bars represent 95% confidence intervals (CIs). A grouped bar chart comparing Narrix and the baseline on two categories: (A) learning narrative strategies and (B) remixing narrative strategies. Panel A has three sub-metrics: (A.1) quantity of retained strategies, (A.2) quality of retained strategies, and (A.3) confidence in understanding strategies. Panel B has three sub-metrics: (B.1) number of strategies applied, (B.2) confidence in applying strategies, and (B.3) satisfaction with applying strategies. For every metric, Narrix shows higher mean values than the baseline (e.g., Narrix participants retained more strategies, reported higher confidence, and applied more strategies). Error bars represent 95% confidence intervals (CIs). Statistical test results above each bar pair (t-tests or Wilcoxon tests) show significant differences in all comparisons, favoring Narrix. Overall, the figure indicates that Narrix substantially outperforms the baseline in both learning and remixing narrative strategies.
| Plot | Character | Information | Emotional | Linguistic | Pacing | Thematic | Engagement | Total | |
|---|---|---|---|---|---|---|---|---|---|
| # Tracks | 9 (23.68%) | 6 (15.79%) | 3 (7.89%) | 10 (26.32%) | 3 (7.89%) | 3 (7.89%) | 1 (2.63%) | 3 (7.89%) | 38 |
| # Strategies | 29 (30.21%) | 19 (19.79%) | 5 (5.21%) | 18 (18.75%) | 10 (10.42%) | 5 (5.21%) | 0 (0.00%) | 10 (10.42%) | 96 |
Table 2. Usage of tracks across creative dimensions in the Remixer workspace of Narrix . Row 1 shows the number of tracks per dimension added by all participants; Row 2 shows the number of distinct strategies added within those tracks from all participants.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Turning Point | Description |
|---|---|
| Opportunity | The introductory event that sets the stage for the narrative. |
| Change of Plans | A pivotal moment where the main goal of the narrative is defined or altered. |
| Point of No Return | The commitment point beyond which the protagonists are invested in goals. |
| Major Setback | A critical juncture where the protagonists face significant challenges or failures. |
| Climax | The peak of the narrative arc, encompassing the resolution of the central conflict. |
Table 3. Definition of turning points ( Tian et al., 2024 ) .
| Dimension | Description |
|---|---|
| Plot | Strategies for plot construction and story progression, e.g., causation, escalation, conflict setup and resolution, reversals and twists, and act and beat frameworks. |
| Character | Strategies for character development and portrayal, e.g., growth, traits, relationships, and archetypal roles. |
| Information | Strategies for information control and perspective, e.g., revelation, concealment, misdirection, foreshadowing, and point-of-view manipulation. |
| Emotional | Strategies for emotional effect, e.g., tension, empathy, surprise, catharsis, and atmosphere. |
| Linguistic | Strategies for language style, e.g., voice, imagery, syntax and rhythm, dialogue, and rhetorical devices. |
| Pacing | Strategies for pacing at the moment and segment levels, e.g., scene versus summary, time compression and expansion, beat density, sentence and paragraph cadence, cutaways and cross-cutting, time skips, and arrive-late, leave-early trims. |
Table 4. Definition of creative dimensions ( Bal, 2004 ; Kennedy and Gioia, 2016 ; McKee, 1997 ; Prince, 2012 ) .
| Story Excerpts | Extracted Strategies & Explanations |
|---|---|
| “ …Maya and Willy went to the glow worms, but they took Maya and Willy up to wreck the hive and the precious sunstone, much to the displeasure of the Queen… ” | Use of MacGuffin : The “precious sunstone” functions as a MacGuffin, an object that propels the plot and is central to the characters’ actions and the conflict. |
| “ …Will reaches a lower road and encounters a man who turns out to be a Frank’s guard, but talks him into letting him pass. He finds Frank’s camp in the woods, a derelict farm, the main building of which is an extensive but rudimentary drug lab. Most of the inhabitants are unconcerned at his presence… ” | Unexpected Lack of Conflict : Rather than a tense or violent encounter, Will persuades the guard and the camp’s inhabitants are unconcerned, subverting expectations for dramatic confrontation and promoting intrigue about their motives or broader circumstances. |
| “ …Reconstructing K.C.’s consciousness, the Moon’s operating system appears to him as his cat, Fuzz Aldrin, and his mother, remarking that they must now ‘get started’… ” | Abrupt Narrative Start : The action begins in medias res, with K.C.’s consciousness being reconstructed and immediate engagement with the central mystery. This technique hooks the reader and leaves important details to be filled in later. |
| “ …After Lydell is arrested in a DEA sting, a desperate Liza asks Neel for help paying for the surgery but he exhorts her to use Phoebe’s situation as ‘fire’ for inspiration, like he did when his wife was dying from cancer and he came up with Lonafen. When a friend’s husband dies from a Lonafen overdose and her condolences are rejected, Liza agrees to testify to the U.S. Attorney’s office that’s investigating Zanna. She admits her involvement in Zanna’s speaker and bribery programs… ” | Chain Reaction Causality : The plot unfolds through a tight sequence of cause and effect, where each character’s action directly provokes consequential reactions, propelling the narrative forward and intensifying momentum. |
| “ …Meanwhile at the hive, Crawley tries to fix the sunstone, but to no avail, much to his dismay. Miss Cassandra calls out that Maya and Willy has left the meadow… ” | Direct Character Dialogue : Involving Miss Cassandra’s speech adds immediacy and a sense of presence to the scene, making the action feel ongoing and interactive while also revealing plot developments through character voices. |
| “ …Alone at last, Shirley allows Stanley to read her work on Hangsaman. He declares it to be a work of genius. Shirley acknowledges his praise. It’s obviously a familiar pattern in their peculiar partnership. The two celebrate by drinking and dancing — together and yet alone in their cluttered house… ” | Symbolic Setting : The ’cluttered house’ setting at the conclusion operates symbolically, reflecting the characters’ internal lives and the complexity of their relationship. |
Table 5. Failure cases in LLM-extracted strategies and explanations flagged by raters.
| Story Excerpts | Extracted Strategies & Explanations |
|---|---|
| “ …Maya and Willy went to the glow worms, but they took Maya and Willy up to wreck the hive and the precious sunstone, much to the displeasure of the Queen… ” | Use of MacGuffin : The “precious sunstone” functions as a MacGuffin, an object that propels the plot and is central to the characters’ actions and the conflict. Lexical Cues : “ the precious sunstone ” |
| “ …Reconstructing K.C.’s consciousness, the Moon’s operating system appears to him as his cat, Fuzz Aldrin, and his mother, remarking that they must now ‘get started’… ” | Abrupt Narrative Start : The action begins in medias res, with K.C.’s consciousness being reconstructed and immediate engagement with the central mystery. This technique hooks the reader and leaves important details to be filled in later. Lexical Cues : “ Reconstructing K.C.’s consciousness ” |
| “ …After Lydell is arrested in a DEA sting, a desperate Liza asks Neel for help paying for the surgery but he exhorts her to use Phoebe’s situation as ‘fire’ for inspiration, like he did when his wife was dying from cancer and he came up with Lonafen. When a friend’s husband dies from a Lonafen overdose and her condolences are rejected, Liza agrees to testify to the U.S. Attorney’s office that’s investigating Zanna. She admits her involvement in Zanna’s speaker and bribery programs… ” | Chain Reaction Causality : The plot unfolds through a tight sequence of cause and effect, where each character’s action directly provokes consequential reactions, propelling the narrative forward and intensifying momentum. Lexical Cues : “ After Lydell is arrested in a DEA sting, a desperate Liza asks Neel for help ”, “ When a friend’s husband dies from a Lonafen overdose and her condolences are rejected, Liza agrees to testify ” |
| “ …Hutch Mansell sits inside of an interrogation room, heavily bruised and injured. Two FBI agents interrogate him about his identity, before Hutch responds that he’s simply "nobody." In a flashback to a few days earlier, Hutch leads an ordinary, mundane life as an office worker with his emotionally estranged wife Becca, teenager Blake and a much younger daughter, Abby… ” | Emotionally Distant Familial Relationships : By describing Hutch’s home life as ’emotionally estranged,’ the story hints at inner conflicts and character motivations, adding depth and potential for emotional development. Lexical Cues : “ with his emotionally estranged wife Becca ” |
| “ …Meanwhile at the hive, Crawley tries to fix the sunstone, but to no avail, much to his dismay. Miss Cassandra calls out that Maya and Willy has left the meadow… ” | Direct Character Dialogue : Involving Miss Cassandra’s speech adds immediacy and a sense of presence to the scene, making the action feel ongoing and interactive while also revealing plot developments through character voices. Lexical Cues : “ Miss Cassandra calls out that Maya and Willy has left the meadow ” |
| “ …Her brother Kevin is caught dancing in the restroom by some older students who record him and bully him, but Clancy’s and Kevin’s mother Margot, the lunch monitor, scares them away. The bullies upload a remix video of Kevin dancing and Margot admonishing them to YouTube, where it quickly goes viral, gaining over two million views… ” | Authority Figure Intervention : Margot, as a lunch monitor and mother, intervenes to protect her son, adding a layer of authority and adult presence that both disrupts the bullying and sets up later plot developments. This juxtaposition of child vulnerability and adult protection intensifies reader investment in the outcome. Lexical Cues : “ Margot, the lunch monitor, scares them away ” |
Table 6. Failure cases in LLM-extracted lexical cues flagged by raters.
| ID | Gender | Age | Writing* | English** | Example*** | AI Writing Tools | AI Usage Freq. |
|---|---|---|---|---|---|---|---|
| P01 | Female | 27 | 1 | 4 | Always | ChatGPT, Grammarly, DeepL | Weekly |
| P02 | Male | 28 | 2 | 5 | Always | ChatGPT, Grammarly, Cursor | Daily |
| P03 | Male | 25 | 3 | 5 | Sometimes | ChatGPT, Cursor | Daily |
| P04 | Female | 25 | 2 | 6 | Often | ChatGPT, Grammarly | Daily |
| P05 | Female | 28 | 1 | 5 | Always | ChatGPT | Weekly |
| P06 | Male | 25 | 2 | 5 | Always | ChatGPT, Grok, Gemini, Grammarly | Daily |
Table 7. Demographic information of participants and their usage of examples and AI writing tools in daily writing practice. *Writing proficiency; **English proficiency (both on Likert scales from 1/low to 7/high); ***Frequency of referring to others’ work in writing (from Never to Always).
Figure 10. Baseline system interface, which includes the same Markdown editor as Narrix , example stories presented as whole-text cards without surfacing narrative strategies, and a chat-based AI assistant, similar to interfaces such as ChatGPT Canvas. Baseline system interface, which retains the same Markdown editor as Narrix but removes the interactive story arc visualization. The Browser panel displays example stories as whole-text cards without surfacing narrative strategies. The Remix panel is replaced with a chat-based AI assistant, allowing users to send free-form prompts and receive conversational responses, similar to interfaces like ChatGPT Canvas.