TL;DR
Inkitiate is a spatial knowledge workspace that allows users to initiate and steer AI agent assistance using digital ink—handwriting and sketching—directly on a canvas. By combining linguistic requests with spatial and graphical marks, users can retrieve, analyze, and organize information in a persistent environment where AI results become immediate building blocks for further thought.
What Problem Does This Paper Address?
Knowledge work is rarely a straight line; it’s a "sensemaking" process of gathering, interpreting, and organizing information. While Large Language Models (LLMs) are powerful assistants, current interfaces often force a "chat-box" paradigm that is disconnected from the spatial way people naturally think (e.g., sketching on whiteboards or annotating margins).
Existing AI tools often lack context about the user's current workspace and provide "terminal" results—static text blocks that are hard to reuse. The authors identify a need for a "pen-first" interaction model where ink acts as both content and instruction, allowing users to steer agents through the same medium they use to externalize their own ideas.
Methods
The researchers developed Inkitiate, a web-based spatial workspace optimized for stylus-enabled devices. The system's design was informed by a formative study with eight participants and is built on four core design goals:
- Task Specification: Combining ink with existing canvas materials to define tasks.
- Bounded Context: Carrying forward relevant history without cluttering the agent's view with the entire canvas.
- Inspectability: Keeping agent execution transparent and interruptible.
- Persistent Results: Writing results back as editable, reusable canvas objects.
Key technical features include:
- Bounded Task Input: Users "lasso" ink and objects to define the scope of a query.
- Handwritten Sigils: Special marks like
#(to bind resources) and//(to specify agent capabilities like "Find Resource") provide explicit control. - Agent Activity Cards: These cards appear on the canvas to show the agent's interpreted intent, current progress, and elapsed time, allowing users to stop or steer the task mid-execution.
- Architecture: A host server mediates between the persistent canvas and a general-purpose agent runtime (using GPT-5.6-Sol), ensuring that AI actions are bounded and results are structured before being written back.
Datasets
As a system-focused paper, Inkitiate was not trained on a new dataset. Instead, the authors performed a Scenario-Based System Validation using 34 distinct ink-query cases to ensure stability and correctness across various combinations of ink and object types. The primary evaluation was a user study with 18 participants who performed both prescribed tasks (Reproduction) and open-ended exploration (Freeform Creation).
Key Insights and Ideas
- Ink is Multimodal: The study identified five distinct roles for ink: Linguistic (writing), Depictive (sketching), Referential (pointing to objects), Relational (connecting ideas), and Spatial (layout cues). Users rarely used just one; they typically combined them to create expressive, compact requests.
- Prior Outputs are Materials: In the freeform study, 69.1% of queries reused previously generated results. This confirms that in a spatial workspace, AI results are not the end of a conversation but the beginning of the next step.
- Bounded Continuity: Users don't want the agent to see the whole canvas, but they do want it to remember the "thread" of the current task. Inkitiate achieves this by defining a "neighborhood" of context based on the location of the ink query.
- The Value of Transparency: Participants highly valued the ability to see how the agent interpreted their messy handwriting before it finished a long task, allowing them to intervene early if the agent went off-track.
Key Results
- High Usability: The system received an average System Usability Scale (SUS) score of 5.93 out of 7, with particularly high marks for learnability (6.33) and user confidence (6.28).
- Low Workload: NASA-TLX ratings were consistently below the scale midpoint, suggesting that steering an agent with ink is not mentally or physically taxing.
- Complex Structuring: Users successfully used the agent to transform raw information into complex visual structures like timelines, branching trees, and radial constellations.
- Steerability: The "Agent Activity Card" proved effective; for example, one user stopped an agent that was trying to generate a map and redirected it to retrieve a real-world map using the
//Find Resourcesigil.
Limitations and Future Work
- Fine-Grained Targeting: While the system handles whole objects well, it struggles when users point to specific words inside a text block or specific regions inside an image.
- Interpretation Uncertainty: Free-form ink is inherently ambiguous. While the clarification interface helps, future versions could benefit from better handwriting and vision models to reduce the need for manual correction.
- Long-term Use: The study was limited to single sessions. The authors suggest that longitudinal studies are needed to see how personal "ink vocabularies" evolve over weeks or months of daily work.
- Multimodal Integration: Participants expressed interest in combining ink with speech for longer, more detailed instructions, suggesting a path for future "sketch-and-speech" interfaces.
Summary updated