TASKARTISAN: Designing Composable Generative Widgets for LLM-Assisted Analysis

While LLM-powered chatbots have become standard tools for data analysis, they often treat each session as a static, isolated event. Users frequently struggle to reuse successful workflows or navigate long, cluttered chat histories. Recent advancements allow LLMs to generate functional Graphical User Interfaces (GUIs), but these are typically "ephemeral"—tightly coupled to specific data and difficult to adapt for future tasks. This paper explores how to transition from these one-off generative UIs to reusable, composable analysis tools.

The authors developed TaskArtisan, a technology probe designed to augment generative UIs with modularity. Unlike standard chatbot artifacts, TaskArtisan decouples three core elements: the data input, the LLM-based analysis logic, and the visual interface component. This architecture allows users to create individual "widgets" that can be assembled into sequential or fan-out pipelines. By providing an infinite canvas, the system enables users to build custom dashboards where widgets can be independently modified, reused with new datasets, and connected to form complex, repeatable analysis workflows.

A controlled study with 12 professionals compared TaskArtisan to standard generative UI tools (Claude Artifacts). While participants spent more time initially authoring workflows in TaskArtisan, they completed subsequent, repeated analysis tasks significantly faster. The study highlights a fundamental trade-off: modularity increases the cognitive load of initial specification but provides long-term efficiency gains.

To guide future development, the authors propose a three-axis design framework:

• Malleability: How easily a UI can be reconfigured or extended after generation.
• Specification: The degree of explicit instruction required from the user to define widget behavior.
• Interoperability: How effectively a UI connects to and exchanges data with other components.