TaskArtisan: Designing Composable Generative Widgets for LLM-Assisted Analysis

Abstract—People increasingly use chatbots such as ChatGPT for everyday analysis tasks. While chatbots unify many analysis functions (e.g., scripts, visualizations, summaries), long conversations become hard to navigate, making it difficult to revisit prior steps or reuse successful workflows. LLMs now generate high-fidelity GUI code that enables people to create customized analysis tools beyond text. Yet, what new opportunities generative UIs bring to analysis work remain unclear. We interviewed six professionals about analysis with chatbots, analyzed publicly shared LLM-generated GUI tools, and conducted a comparison study (N=12) between a chatbot and TASKARTISAN, a technology probe that enables people to create and assemble generative analysis UI widgets for sequential and fan-out composition. We find that GUI improved clarity and visual presentation but also introduced rigidity and additional prompting challenges. We summarize the trade-offs into a provisional design framework (malleability, specification, interoperability) to inform future generative UI in LLM-assisted analysis workflows. Index Terms—Generative UI, Large language models, Modularity, Design

Introduction. People analyze data (e.g., text, spreadsheets) [1], [2] to surface results (e.g., summaries, visualizations) and ultimately make decisions (e.g., select a college, plan inventory) [3]– [5]. Traditionally, people use multiple applications to complete such analyses [3], [6]. For example, a college counselor might use a browser to look up school information, then use Excel to plot expected cost against financial aid, and finally use Google Docs to present the pros, cons, and images of each school in a table. New Large Language Model (LLM) chatbots like Chat- GPT Analyst [7] and Claude Artifact [8] unify diverse analysis and visualization capabilities into one application. Using these chatbots, people can upload their data then write a prompt to describe what analysis they want to perform. Thus, they can complete an entire analysis workflow in a single application (e.g., create a plot then a comparison table).

Discussion / Conclusion. Based on our findings, we propose a three-axis provisional design framework to inform future generative UI design in LLM-assisted analysis (Figure 1). Each axis reflects a trade-off observed in our study. Low vs. High Malleability. The malleability dimension describes how easily a generative UI can be customized and extended after creation, ranging from a fixed artifact to one supporting user-defined reshaping of both UI and analysis logic. Low-malleability UIs suit stable, well-defined tasks, but create friction when analysis needs evolve mid-workflow, which our participants frequently encountered. At the malleable end, users can adjust analysis granularity, modify UI structure, and recompose workflow steps, but our study found this increases authoring overhead: non-programmers (P4, P9) struggled to specify widgets at the right granularity and needed to “think like a software engineer.” Future systems should let users define fixed and customizable parts of UIs [15] or enable low-friction manipulations via natural language [54]. Implicit vs. Explicit Specification.