01. The Problem
Generative AI tools like ChatGPT, Claude, and Gemini have become everyday tools for students. But despite their expanding capabilities, they all share the same standardized chat interface. They are designed for conversation, not for the complex, multi-threaded, iterative way students actually work.
Typical ChatBot Wireframe
The left panel contains past conversations. This inefficiently takes up valuable screen space for a feature that does not need to be so quickly accessed. The right panel has the active conversation which is the main feature in use.
Responses and outputs progress linearly and vertically. Responses go up in the conversation and consequently get lost under long outputs. If a response was more than 3-4 prompts ago, consider it lost. Only one conversation or project is in use at a time.
Can this be designed better for students?
02. Research Process
Competitor review of common and student-oriented AI tools, followed by 5 user interviews with students across diverse majors and AI usage levels. Interviews included an AI use-case walkthrough. Findings were synthesized through affinity mapping into 3 personas and a journey map.
Methods:
-
5 user interviews
-
Competitor Review
Competitor Review between conversation-based models (e.g. ChatGPT, Claude, Gemini) and prototyping-based models (e.g. v0 by Vercel, Lovable, Codex)
Affinity Map from user interviews
Personas & Journeymap:
03. Discovery Insights
1. Chat iteration breaks continuity
Users working across multiple projects lose prior outputs as sessions grow. Information gets buried and can’t be recovered without re-prompting. To solve this problem, the interface can pin responses to a workspace panel, where notes and relevant information can be saved, while chats can become lost.
2. Users want control over parts, not wholes
High-control users prefer to do fine edits themselves. Full regeneration feels like a loss of ownership over their work. A solution to this problem is an interface that allows users to freely edit, or spot edit, the output while remaining in the AI system for repurposing.
3. Prompting is heavy and risks information loss
The effort of providing context slows users down. The more they prompt, the more nuance gets dropped from the output. A solution to this problem includes a simultaneous task panel that allows users to work on multiple problems with the same conversation, without getting lost and without having to reprompt in a new conversation.
04. Prototype
Consolidated 3 design directions for prototyping.
1. Centralized Product Development Space
2. Spot-editing and option-based suggestions
3. Create Simultaneous Task Panel
Multiple tabs within a single conversation.
This allows for conversations to span multiple projects, to retain memory between different projects and update them simultaneously.
Conversation Panel.
This allows the messages to be short, and also be added to the workspace panel to avoid information loss. It also centralizes the conversation space.
Editable Workspace Panel.
This allows for the focus to be the work and output, not lengthy conversation. It also allows for users to edit the work that the AI outputs, based on its suggestions.
*Created AI mockup with v0 and Codex
05. Usability Testing
-
5 students usability tested
-
2 tasks
-
The first task was to task-switch between a paper and a presentation in a single conversation.
-
Users would begin with writing a paper, then prompt the model to create a presentation based on the paper. Then after making edits to the presentation, the paper should be updated accordingly. This workflow tested the usability of the simultaneous tabbing and the centralized work panel, which found that users appreciated one workspace for multiple branching and returning tasks without the need to re-prompt the model.
-
-
The second task was to spot edit an essay.
-
Users would be asked to revise the hook of an output from the model, and replace a singular word. There were several ways this could be done, which included highlighting the section, or asking the chatbot directly. The model would output suggestions which could be "applied" or "discarded". This workflow tested the ability to take autonomous control over model outputs.
-
-
Key findings:
Overall, 5 out of 5 tested students preferred the simultaneous tab implementation, noting that it directly reduced information loss.
Additional Usability Findings:
Users preferred to preserve user autonomy whenever possible. Users also wanted spot edits to consider full document context, not operate in isolation. So, suggested edits by the model (i.e. "apply" or "discard") were strongly preferred over forced replacement and direct changes by the system. Likewise, systems like these should allow students to control the final outputs themselves, since they contain more holistic context.