Users can input hand-drawn sketches, text, and reference images simultaneously on the same canvas; the system understands ambiguity and emotion. Then, they create precise versions and three-way neighborhood divergence with one click, allowing for instant comparison and selection.
Users can semantically lock text and image, and change one side while the other updates in real time (linking style, tone, and camera language). Lasso for localized deepening, Collage for spatial collage, Filter for style/tone linkage, and Perspective for switching camera position and narrative perspective.
Cluster scattered cards into a "story skeleton" by character, scene, or theme. Export a storyboard, script, and shot list with one click..
Designing an AI-Native Narrative Tool for Fiction Writers.
Successfully published a research paper at CHI 2026, establishing a novel theoretical framework for AI-assisted narrative that integrates dual coding theory and instrumental interaction.
Achieved a 40% reduction in context-switching and a 35% increase in Cognitive Offloading scores through the implementation of the Canvas and Cluster system.
Engineered a node-based interaction model that led to 80% of writers reporting "serendipitous discoveries" during their creative process.
When this project started, I didn't open a blank canvas. Instead I opened every competing tool I could find. What I discovered wasn't inspiration. It was a pattern: every existing product had quietly agreed that text and image belong in separate workflows. Writing tools made text smarter. Generation tools made images richer. But none had questioned whether that separation was the right premise at all. I stopped there, named the gap, and made it the actual starting point of the design. That reframe shaped everything that came after. Research isn't just collection — it's interrogation.
As discovery progressed into design, the team split into two directions: one prioritized innovation through a canvas-first system; the other prioritized usability through a pain-point-driven approach. Neither was wrong, but together they created a deadlock. I didn't pick a side. Instead, I proposed an adjustable panel structure, something writers could resize or collapse depending on where they were in their process. When engineers pushed back on interaction complexity, I didn't defend the design from instinct. I pulled out interview findings and research precedents and built the case from evidence. We reached a consensus.
Once the direction became clear, I assumed the hardest part was behind us,but it wasn't. During handoff, engineers struggled to understand how the Lasso tool's selection logic connected to the AI's semantic extraction process. One suggested replacing the whole multimodal interaction with a standard chat box. I wrote detailed Interaction Specs: pixel-level annotations, data flow diagrams, and explicit rationale for every decision that might get simplified away. Not because I didn't trust the team — but because I'd learned that intent doesn't survive a handoff on its own. Clarity has to be built into the document itself.
