Myna AI
| 2025
Designing Interfeaces for AI That Executes Work
TL;DR
Myna operated inside a multimodal chat interface. Users could generate responses, create social posts, and analyze content, but as tasks grew complex, chat alone couldn't manage them. I designed a task card system that bridged conversation and execution, then scaled it into a reusable foundation for future AI workflows.
Role
Product Designer
Impact
Structured AI Workflows
Introduced task cards that turned chat responses into clear, reviewable units of work.
Faster Decision-Making
Users could quickly scan outputs, edit, and approve actions without digging through long chat messages, guide or correct the system when needed and track work as it progressed
Reusable Interaction Model
The task card became a scalable pattern used across multiple workflows like review responses and social media posting.
Foundation for Future Workflows
Modular UI primitives enabled the system to support new AI tasks without redesigning the interface.
The Problem
Chat works well to start work. It breaks down when managing it.
Myna's chat interface handled simple requests well. But as tasks grew more complex — multi-step, parallel, requiring review — the format collapsed.
Messages made it hard to track ongoing work, review outputs quickly, or manage tasks across steps. Everything lived in the thread with no stable structure to act on.
Chat was the right place to initiate work. It wasn't built to manage it.
The Guiding Decision
Don't replace chat. Anchor work inside it.
Instead of building a separate task management layer, I introduced task cards directly inside the conversation thread. Each card represents a single unit of AI work — triggered from chat, reviewed in place, acted on immediately.
This creates a clear interaction loop:
Conversation → Task Card → Review → Decision
Designing the Task Card
One card. Three layers. A complete unit of work.
The task card became the bridge between conversation and execution. Every card carried three things:
Context
Why the task exists: the original prompt, uploaded file, or user request.
Status / Output
The task lifecycle made visible: generating → ready for review → completed. Structured results in place of long chat responses.
Decision layer
Clear actions at every stage: Approve, Edit, or Regenerate.
This structure separated conversation from execution — making each piece of work easy to find, review, and move forward without scrolling through the thread.
Human-in-the-Loop Interaction
Each card surfaces clear moments where users guide or validate the system.
This approach kept the natural flow of conversation, while giving each piece of work a stable place in the interface.
Scaling the Pattern
One card proved the model. Primitives made it scalable.
Once the task card worked for a single workflow, the structure revealed something more useful: the card wasn't tied to any specific task. It was a container for AI work inside the conversation.
The same pattern could represent a review response, a social media post, a document summary, or a content analysis. What mattered wasn't the task, it was the interaction pattern underneath it:
Context → Status → AI Output → Decision
To scale this across workflows, I broke the card into reusable UI primitives:
Progress indicators
Feedback modules
Action triggers
Together, these primitives let new AI workflows be assembled inside chat without redesigning the interface each time.
Trade-offs Made
Chose structure and scalability over flexibility and speed.
Traded away
Freeform chat flexibility
Per-workflow UI design
Full automation
In favor of
Structured, reviewable outputs
Reusable primitives across all workflows
Human review at every decision point
Key Takeaways
Designing for AI inside chat changes what an interface needs to do.
1
Start with conversation
Chat is the most natural way for users to express intent. It became the entry point for every task, not just simple ones
2
Add structure where work happens
AI outputs need stable UI containers so users can review, correct, and move work forward
3
Build patterns, not screens
Breaking the task card into primitives meant new workflows could be assembled without starting from scratch
Next Steps
1
Progressive Onboarding
Introduce the interaction model gradually. Users start with simple chat tasks, then unlock editing outputs, correcting results, and triggering follow-up steps.
2
Long-Running Campaigns
Extend the system to support ongoing work: tracking multiple task cards, monitoring progress, and managing outputs over time.