Abstract
Effective collaboration with generative AI systems requires users to clearly communicate their intents (intent-based outcome specification). Yet such intents are often underspecified and evolve during interaction, dynamic support for intent communication is essential. Through a systematic literature review of 33 papers, we synthesize a structured understanding of intent communication, identifying four key aspects: articulation, exploration, management, and synchronization. Building on these findings, we derived design implications that translate them into actionable design and implemented IntentFlow, a system for LLM-based writing that realizes these implications through adjustable UIs, intent-to-output linking, and versioned refinement. A technical evaluation (N=60) and a within-subjects study (N=12) confirm that IntentFlow helps users discover, elaborate, and consolidate their intents into a curated set. Interaction logs further reveal a shift from reactive error correction to proactive intent refinement. Our work demonstrates how a system effectively designed to support these four communication aspects can substantially enhance human-LLM interaction.
Introduction
In the current paradigm of Generative AI, interaction has shifted toward intent-based outcome specification. Users no longer provide step-by-step instructions but instead specify the desired end state. However, human intent is rarely static or well-defined at the outset; it is often subconscious, underspecified, and fluid. Standard chat-based interfaces struggle with this fluidity, often trapping users in repetitive "prompt engineering" loops where they attempt to correct misaligned outputs by reformulating text in a black-box environment.
The Four Pillars of Intent Communication
To address these challenges, we conducted a systematic literature review of 46 HCI papers to identify the core requirements for supporting intent. We synthesized these into four pillars:
- Articulation: Assisting users in translating vague, high-level ideas into concrete, actionable instructions.
- Exploration: Encouraging users to discover alternative directions or "what-if" scenarios through structured feedback.
- Management: Providing tools to organize, track, and persist intents as they evolve throughout a session.
- Synchronization: Establishing a transparent, bidirectional link between specific user intents and the corresponding segments of the AI-generated output.
The IntentFlow System
Building on these pillars, we developed IntentFlow, a writing assistant designed to move beyond the simple prompt box. The system reifies intents as first-class UI elements within a dedicated Intent Panel.
When a user provides a vague initial input, the system extracts a high-level goal (e.g., "Write a scientific article") and a list of low-level intents. These intents are further decomposed into Intent Dimensions—interactive controls such as sliders for length, radio buttons for focus, and tags for tone. This structure prevents intents from being lost in chat history and allows for precise, granular control over the output.
A critical feature of IntentFlow is bidirectional traceability. By hovering over an intent in the panel, the system highlights the specific text in the output panel influenced by that intent. This "glass box" approach allows users to verify alignment and identify gaps in the AI's understanding, often triggering the discovery of new, necessary intents.
Evaluation: From Correction to Proactive Refinement
We evaluated IntentFlow through a within-subjects study (N=12) comparing it to a standard chat baseline. Behavioral analysis revealed a significant shift in user interaction patterns:
- Behavioral Shift: Users in the baseline spent the majority of their time in "breakdown-recovery" cycles, repeatedly using Correct actions to fix AI errors (M=4.33). In contrast, IntentFlow users shifted toward Adjust actions (M=4.50 vs M=1.17 in baseline), proactively refining their own intents rather than reactively fixing output.
- Reduced Cognitive Load: NASA-TLX scores indicated a significant reduction in perceived effort and frustration. Participants felt they had a more effective "steering wheel" for the model, leading to higher ratings for intent alignment.
- Transition Stability: Markov chain analysis showed that IntentFlow interactions stabilized around refinement, whereas baseline interactions were characterized by volatile cycles of error and rollback.
Conclusion
IntentFlow demonstrates that by treating intent as a dynamic process rather than a static command, we can significantly improve the human-LLM collaborative experience. By providing structured tools for articulation, exploration, management, and synchronization, systems can help users "debug" their own thoughts and achieve higher-quality outcomes with less frustration. Future work will explore the persistence of these structured intents as long-term interaction artifacts across different domains and tasks.
Summary updated