ADR-0007: Conversation-First UX¶
Status: Accepted Date: 2026-03-26 Session: 17/22
Context¶
Curaway's original UI followed a traditional multi-page CRUD pattern: register on one page, upload documents on another, view matches on a third, manage consent on a fourth. Each step had its own form, its own navigation, and its own mental model.
Early user testing (Session 17) revealed significant friction:
- Patients dropped off between steps, especially at the document upload page.
- The multi-page flow felt clinical and bureaucratic -- exactly the experience patients are trying to escape.
- Users did not understand the relationship between uploading documents and receiving matches.
- The step-by-step structure forced a linear workflow that did not match how patients naturally think about their healthcare journey.
Decision¶
Pivot to a conversation-first UX inspired by Claude.ai and ChatGPT. A single chat interface drives the entire patient journey: onboarding, document upload, record review, matching, and consent are all handled through natural conversation. The agent orchestrator routes user intent to specialized sub-agents transparently.
Rationale¶
- Reduced friction. One interface, one interaction model. Patients type or upload, and the system guides them. No navigation, no forms, no "which page am I on?" confusion.
- Natural interaction. Patients are already comfortable with chat interfaces (iMessage, WhatsApp). A conversational interface feels familiar and approachable, reducing the intimidation factor of medical workflows.
- Guided discovery. The agent can proactively guide patients: "I see you uploaded a pathology report. Would you like me to extract your diagnosis and look for matching trials?" This is impossible in a static multi-page flow.
- Flexible ordering. Patients can upload documents, ask questions, and review matches in any order. The conversation adapts. The CRUD flow forced a rigid sequence.
- Agent orchestration. Behind the chat interface, an orchestrator agent routes messages to sub-agents (intake agent, document agent, matching agent, consent agent) based on intent classification. The patient sees a single coherent conversation.
Alternatives Considered¶
| Alternative | Pros | Cons | Verdict |
|---|---|---|---|
| Keep multi-page CRUD | Proven pattern, familiar to developers | High drop-off rate, intimidating for patients, rigid workflow | Rejected after user testing |
| Wizard-style steps | Clearer progress indication, guided flow | Still form-based, still linear, still intimidating | Rejected |
| Hybrid (chat + dashboard) | Chat for interaction, dashboard for review | Two mental models, complex navigation, split attention | Planned for power users (v2) |
Consequences¶
- Positive: Dramatically lower friction for patient onboarding. The chat interface is the only thing patients need to learn.
- Positive: Agent orchestration enables sophisticated multi-step workflows without exposing complexity to the user.
- Positive: Conversation history provides a natural audit trail of patient interactions.
- Negative: Building a reliable agent orchestrator is significantly more complex than building CRUD pages. Intent classification must be accurate, and sub-agent handoffs must be seamless.
- Negative: Some information is harder to present in a chat format (tables of matches, document summaries, consent forms). Rich message types (cards, carousels, inline forms) are needed.
- Negative: Accessibility and screen-reader support for a chat interface requires careful attention.
- Accepted risk: If the agent misclassifies intent, the patient may be confused. Mitigated by a "I'm not sure what you mean" fallback and the ability to explicitly request actions ("show my matches", "upload a document").