Intent Tagging: Exploring Micro-Prompting Interactions for Supporting Granular Human-GenAI Co-Creation Workflows

Generative artificial intelligence (GenAI) offers substantial potential for automating and assisting complex content creation tasks, such as authoring slide decks. However, users frequently struggle to integrate current GenAI tools into their creative workflows due to three core friction points: difficulty communicating intent accurately through free-form text prompts, rigid linear workflows imposed by chat interfaces, and the challenge of discovering and articulating creative requirements upfront. The article evaluates a novel interaction method called intent tagging—a graphical micro-prompting approach using atomic conceptual units organized on a visual canvas—and demonstrates its effectiveness in supporting flexible, non-linear human-AI co-creation.

To evaluate this approach, the researchers built IntentTagger, a prototype system driven by large language models that enables users to steer presentation generation across narrative, visual style, and content sources using interactive tags, adaptive widgets, and pre-computed real-time previews. The authors conducted a comparative laboratory user study with 12 professional knowledge workers from a major technology company. Participants completed both structured comparative tasks—benchmarking IntentTagger against commercial chat- and design gallery-based tools (specifically PowerPoint Copilot and Designer)—and open-ended presentation authoring tasks, followed by qualitative video analysis and post-task surveys.

The evaluation revealed several critical findings. First, participants felt significantly more in control when guiding generation toward their intended outcome using intent tags compared to chat-based prompting (a mean difference of 2.67 on a 6-point scale). Second, IntentTagger substantially enhanced iterative refinement, with users rating their ability to provide follow-up information to refine output much higher (a mean difference of 2.83). Third, proactive AI suggestions proved highly valuable rather than intrusive: across open-ended tasks, users incorporated an average of 73.5% of their active tags directly from system suggestions to discover new design directions and overcome creative blocks. Finally, the interface successfully accommodated non-linear workflows, allowing participants to move fluidly between deck-wide parameters, outline adjustments, and single-slide refinements in an average creation time of under nine minutes.

These findings imply that replacing monolithic text prompts and rigid conversational chat streams with modular, graphical micro-prompts significantly reduces user frustration, trial-and-error prompting, and task completion time. Graphical intent elicitation promotes reflection-in-action by functioning as an interactive canvas where the system's dynamic suggestions help users clarify tacit goals. However, as workflows become more non-linear, a key trade-off emerges between high generation flexibility and fine-grained visual stability, as users strongly desire the ability to anchor or lock specific approved elements across iterative generation cycles.

Based on these results, software developers and user experience designers building GenAI tools should transition from pure chat interfaces toward hybrid direct-manipulation interfaces that pair micro-prompt tags with adaptive user interface controls. Development roadmaps should prioritize dynamic pre-computed previews and explicit pinning mechanisms that lock finalized design elements. Furthermore, onboarding experiences should provide scaffolding, such as predefined tag templates, to reduce initial user hesitation around tag naming and placement. Before broad commercial deployment, organizations should conduct longitudinal in-situ pilots to observe how intent tagging performs across diverse corporate domains and extended authoring sessions.

These conclusions are supported by a controlled laboratory study with high statistical significance in user satisfaction metrics. However, readers should note certain limitations: the participant pool was limited to 12 corporate professionals from a single enterprise, tasks were constrained to short slide decks of 6 to 7 slides to maintain sub-15-second generation cycles, and the prototype lacked manual pixel-level formatting tools. Confidence in the usability advantages of intent tagging is strong for rapid initial prototyping, but further empirical study is necessary to validate performance in large-scale document production environments.