✲ ADOBE INTERNSHIP ✲ BEING SHIPPED
✲ ADOBE INTERNSHIP
Designing Adaptive In-Chat Experiences for Adobe Workflows.
Designing Adaptive In-Chat Experiences for Adobe Workflows.
MY ROLE
MY ROLE
As a product design intern, I helped define AI design strategy, built an AI auditing workflow, a component compliance tool, and 4 production code components.
As a product design intern, I helped define AI design strategy, built an AI auditing workflow, a component compliance tool, and 4 production code components.
TOOLS
Figma
Claude Code
Internal Tools
TEAM
TEAM
1 Product Manager
1 Product Manager
4 Product Designers
4 Product Designers
12 Engineers
12 Engineers
TIMELINE
TIMELINE
May - August 2026
May - August 2026
✲ CONTEXT
✲ CONTEXT
✲ CONTEXT
The new AI Integration for Marketers Was Promising But not Fully Realized.
The new AI Integration for Marketers Was Promising But not Fully Realized.
The promise of Workfront MCP
Workfront MCP links Workfront functionality (project management, approvals etc.) with LLMs like Claude so marketers can complete tasks, check projects and more all in chat.
This video explains how MCP works through the lens of Mr. Potato Head as the target user, a project marketing manager.
Initial problem framing
PROBLEM
Only ~200 users regularly used Workfront MCP and growth was slow.
OPPURTUNITY
Improve visibility and ease of setting up the MCP.
✲ FIRST SOLUTION
✲ FIRST SOLUTION
✲ FIRST SOLUTION
Improving Onboarding and Awareness Wasn't Fully Successful.
Improving Onboarding and Awareness Wasn't Fully Successful.
Improving visibility and reducing onboarding friction.
Improving visibility and reducing onboarding friction.
I designed a home screen modal to drive MCP feature discovery and partnered with technical writers to streamline setup documentation.
I designed a home screen modal to drive MCP feature discovery and partnered with technical writers to streamline setup documentation.
The popup I designed
The popup I designed
Improved setup instructions
Improved setup instructions
Workfront MCP
Result: more users joined but churn remained high.
Result: more users joined but churn remained high.
The changes were successful in drawing more users to the platform but ultimately failed in creating signficant sustained growth for the MCP.
The changes were successful in drawing more users to the platform but ultimately failed in creating signficant sustained growth for the MCP.
Reframing
Improving onboarding was just a piece of the puzzle.
To keep users, the MCP itself has to be improved.
✲ SCOPING MCP DESIGN
✲ SCOPING MCP DESIGN
✲ SCOPING MCP DESIGN
Auditing the Existing MCP Experience and Usage Data.
Auditing the Existing MCP Experience and Usage Data.
Auditing with a custom AI agent and usage data.
Auditing with a custom AI agent and usage data.
Analytics, human feedback, and a custom AI prompt-testing agent (shown below) revealed that the existing text-first experience struggled to handle complex actions.
Analytics, human feedback, and a custom AI prompt-testing agent (shown below) revealed that the existing text-first experience struggled to handle complex actions.
Custom agentic workflow
Custom agentic workflow
Grounded prompt generation
AI Prompt injection
Output evaluation
(with human judgement)
Agentic workflow demo
Agentic workflow demo
What I found: designing for LLM's requires nuance
LLM-Native Text output struggles with precision for complex prompts.
Predefined components struggle with latency and customizability for simple prompts.
✲ KEY DESIGN DECISIONS
✲ KEY DESIGN DECISIONS
✲ KEY DESIGN DECISIONS
Deciding component form and
when components surface.
Deciding component form and
when components surface.
Surfacing components only when needed
Defining Logic & AI Integration →
Leveraging my findings I proposed only surfacing components for complex user prompts which my team agreed was the best to create an adaptable experience.
Exploring component forms
efficiency
learnability
Simplicty
Precision
Prioritizing simplicity over precision
Prioritizing simplicity over precision
We prioritized essential functionality, intentionally stripping back traditional design patterns for the MCP to create an escape from existing Workfront's complexity.
We prioritized essential functionality, intentionally stripping back traditional design patterns for the MCP to create an escape from existing Workfront's complexity.
learnability
efficiency
✲ SOLUTION
✲ SOLUTION
✲ SOLUTION
✲ SOLUTION
Centralize Marketing Workflows Inside LLMs.
Centralize Marketing Workflows Inside LLMs.
Centralize Marketing Workflows Inside LLMs.
Centralize Marketing Workflows Inside LLMs.
LLM integration eliminates Workfront friction by matching user intent with text dialogue or custom, action-specific controls.
LLM integration eliminates Workfront friction by matching user intent with text dialogue or custom, action-specific controls.
LLM integration eliminates Workfront friction by matching user intent with text dialogue or custom, action-specific controls.
LLM integration eliminates Workfront friction by matching user intent with text dialogue or custom, action-specific controls.
Components built in the production codebase
Components built in the production codebase
01 Approve an asset
With this component reviewers can approve, deny or add feedback to assets in chat.
02 Remind Approvers
With this component project managers can check approval processes and remind approvers in bulk or individually in chat.
03 Request Approvers
With this component project managers can set up approval chains with multiple steps and customize the process.
04 Discover assets
With this component anyone on the marketing team can view assets in a set project for purposes like sending to colleagues.
Making component design and building scalable.
Making component design and building scalable.
I built an internal tool to help other teams build and design components in the production codebase. It leverages checklists and a sortable gallery view to ensure consistency with the existing components and the design principles I helped define.
I built an internal tool to help other teams build and design components in the production codebase. It leverages checklists and a sortable gallery view to ensure consistency with the existing components and the design principles I helped define.
The Results
The Results
User churn decreased from around 32% to 22% as components began to launch.
The framework I helped develop now influences cross-product AI design strategy at Adobe.
Other teams became enabled to build their own components allowing more to contribute.
Designed by the competition’s smallest design team, our system successfully met all NASA specifications.
Designed by the competition’s smallest design team, our system successfully met all NASA specifications.
Designed by the competition’s smallest design team, our system successfully met all NASA specifications.
Here is me presenting Practility at Iltacon 2025
We even got our logo on "stress brains"!
✲ FINAL THOUGHTS
On Designing in Big Teams and for AI.
Designing in a large organization requires getting everyone (PM's, engineers, marketers, etc.) aligned.
Designing for LLM's requires working and testing in code to account for non-deterministic outputs.
Design teams can be engineering teams in 2026 by actively working in the production codebase!
Thanks for visiting! I go into more detail on my work over calls/interviews. Feel free to reach me here.
Thanks for visiting! I go into more detail on my work over calls/interviews. Feel free to reach me here.
But Wait, There's More!
But wait, there's more!
But wait, there's more!
