A Graphical Interface for Structured Prompting with Large Language Models
| March 2026 | University of California San Diego |
Summary
We present a graphical, tree-based interface for structured prompting with large language models (LLMs) that replaces linear chat interactions with an exploratory prompt graph. Instead of a single conversation thread, users construct a branching structure of prompts where each node represents a reasoning step, sub-question, or refinement direction. A pilot study shows that this interface improves perceived organization, supports exploratory learning, and encourages iterative reasoning over linear prompting.
Motivation
Standard LLM interfaces are linear and text-heavy, making it difficult for users to:
- Track multiple reasoning paths
- Revisit earlier ideas easily
- Understand the global structure of exploration
- Apply structured prompting techniques (e.g., CoT, decomposition) without explicit training
While prompt engineering techniques improve model performance, they remain largely inaccessible to non-expert users. We aim to embed these techniques directly into the interface by making prompting visually structured and navigable.
Contributions
We designed and implemented a node-based LLM interaction system where prompts form a dynamic tree. Specifically:
- Built a prompt graph interface where each node represents an LLM interaction
- Enabled multi-branch generation (3–5 child nodes per prompt) for exploration
- Added re-prompting side nodes for refinement and clarification
- Supported navigation, traversal, and revisitation of prior reasoning paths
- Conducted a pilot usability study (n=4) using think-aloud protocols and semi-structured interviews
- Performed thematic analysis to identify user behavior patterns and usability themes
Stack
- Frontend: React
- Graph UI: React Flow (node-based interface system)
- Backend: Node.js session-based web application
- LLM: Gemini API
- Evaluation: Think-aloud protocol, screen/audio recording, thematic analysis (Braun & Clarke)
Challenge & approach
Problem:
LLM interactions are inherently linear, which limits exploratory reasoning, branching thought processes, and structured prompt refinement. Users also struggle to apply advanced prompting strategies without explicit instruction.
Approach:
We reframed prompting as a graph traversal problem instead of a chat stream. Each prompt becomes a node in a tree, and the system generates structured child nodes representing alternative reasoning paths.
This enables users to:
- Explore multiple reasoning directions in parallel
- Revisit and compare prior nodes
- Iteratively refine prompts through side branches
- Maintain a structured mental model of exploration
Additionally, by limiting context to local node state, we reduce context window overload and improve modular reasoning behavior.
Highlights
- Chains of reasoning visualized: Users reported improved understanding of how ideas connect compared to linear chat interfaces
- Exploratory learning behavior: Branching encouraged brainstorming-style exploration of topics
- Mental model alignment: Participants described the interface as matching their natural thought process
- Iterative prompting loop: Users engaged in refinement cycles rather than one-shot queries
- Revisitation needs: Strong demand for history tracking and node revisit functionality
- Tradeoff identified: Branching improves exploration but can increase cognitive load for unfamiliar topics
