Abstract
This position paper proposes adopting Don Norman’s classic "Seven Stages of Action" as a framework to improve the usability of Large Language Model (LLM) writing assistants. By decomposing human-AI interaction into specific cognitive steps—from goal formation to output evaluation—the authors provide a roadmap for designers to bridge the "gulfs" of execution and evaluation that currently lead to user frustration and tool abandonment.
Introduction: The Usability Gap in LLM Interaction
While Large Language Models (LLMs) like ChatGPT have shown immense potential as writing partners, they suffer from significant usability hurdles. Users often struggle with "prompt engineering," representing a significant Gulf of Execution. Conversely, once a model generates text, users often find it difficult to verify if the output is accurate or trustworthy, creating a Gulf of Evaluation.
Current research often prioritizes the technical performance of models over the end-to-end human experience. Issues such as unpredictable performance, a lack of user ownership over generated text, and the difficulty of "steering" a model toward a specific intent frequently result in tool abandonment. This paper argues for a structured, usability-focused approach to designing these interactions.
The Seven Stages of Action Framework
To address these challenges, the authors apply a foundational Human-Computer Interaction (HCI) theory—Norman’s Seven Stages of Action—to the domain of LLM-based writing. This framework breaks down any action into seven cyclical stages, categorized into Execution and Evaluation phases.
The stages are defined as follows:
- Goal Formation: Defining the high-level objective of the task.
- Plan: Decomposing the goal into meaningful chunks suitable for an LLM.
- Specify: Determining the specific prompt or input to use.
- Perform: Executing the prompt and iteratively updating it.
- Perceive: Forming an initial impression of the model's output.
- Interpret: Understanding why the model generated specific content and how it relates to the prompt.
- Compare: Assessing whether the output satisfies the original goal.
By walking through a hypothetical use case—writing a software tutorial for matplotlib—the authors demonstrate how this framework can guide the design of more intuitive interfaces.
Key Insights and Tool Audit
The application of this framework reveals several critical gaps in current LLM research and development:
- The Execution Bias: Most current tools are heavily weighted toward the Execution Phase (Planning, Specifying, and Performing). There is a notable lack of support for the Evaluation Phase (Perceiving, Interpreting, and Comparing).
- Decomposition as a Design Requirement: Effective interaction requires breaking high-level goals into sub-goals. Writing assistants should explicitly support the "Plan" stage to help users manage complex tasks.
- Context-Dependent Evaluation: The "Compare" stage varies significantly by task. For instance, a coding tutorial requires execution for verification, while research writing requires citation checking. Interfaces must be tailored to these specific evaluative needs.
The authors audited existing LLM research tools using this framework:
- Wordcraft & TaleBrush: These tools primarily support the Specify and Perform stages through meta-prompting and sketch-based inputs.
- PromptChainer: This tool is highlighted for supporting the Evaluation phase by visualizing the output of every step in a prompt chain, aiding in Perception and Interpretation.
- PromptMaker: Focuses on the Execution phase by providing interfaces for rapid prompt prototyping.
Conclusion and Future Directions
Using the Seven Stages of Action helps identify under-represented research areas, specifically alignment to human preferences, explainability, and the interpretability of model outputs.
However, the authors acknowledge that in real-world writing, these stages are often blurred; for example, "perceiving" and "interpreting" often happen simultaneously. Future work should focus on developing flexible designs that emphasize different stages based on user needs and task nature. The next step involves user-centric validation through interviews and participatory design to identify task-specific requirements for each stage of action in different writing contexts.
Summary updated