Hero media placeholder
Work Contribution
Replace with the final end-to-end Work Contribution walkthrough.
Work Contribution
Turning everyday work into review-ready evidence with AI-powered warm starts and contextual feedback
Work Contribution is a year-round record that helps employees document projects and reuse them as evidence during performance reviews. I led three focused enhancements that help employees identify work worth documenting, articulate its leadership impact, and strengthen it with feedback from the people closest to the work.
My role
Design Lead Sole product designer
Timeline
Mar - Jun 2026 Launched in 2026
Team
1 PM 3-4 engineers
Scope
2 shipped / pilot 1 in development
The product captured work, but still depended on employee effort
Work Contribution is a year-round record available to Amazon employees. They can document a project at any stage, then reuse that evidence when managers evaluate impact and performance.
01
The work lived elsewhere
Before Work Contribution, employees tracked projects across personal notes and disconnected documents. Remembering what mattered often became a review-season scramble.
02
A blank form demanded reflection upfront
Even after the core creation flow launched, employees still had to remember what to document, summarize it, and translate their work into Leadership Principles on their own.
03
Self-reported evidence lacked context
Collaborators could be named on a contribution, but their first-hand perspective was not connected to the record or reusable during later talent decisions.
Improve three moments instead of redesigning the whole product
I inherited the product after its manual creation flow had launched. As the sole designer, I focused on three high-friction moments where small interventions could make the evidence more complete and useful.
Capture
Identify work worth documenting
Pre-draft contribution suggestions from existing work signals so employees do not have to begin from memory.
Articulate
Connect evidence to leadership behaviors
Use AI as a warm start for mapping descriptions and work artifacts to relevant Leadership Principles.
Validate
Collect context from collaborators
Turn a collaborator field into a timely feedback opportunity tied to the work itself.
How might we reduce the effort of documenting everyday work while making each contribution more useful during performance reviews?
Replace the blank page with relevant work suggestions
The system pre-drafts three to five potential contributions from an employee's own work activity, including code reviews, tickets, and issues.
01
Start with existing context
Suggestions surface inside Work Contribution, where employees can review a title and description without switching tools or reconstructing the project from memory.
02
Keep creation one click away
Selecting a suggestion creates a draft and carries the proposed title and description into the existing form for review and editing.
03
Preserve employee agency
Suggestions are private, collapsible, editable, and paired with feedback controls. Nothing becomes a completed contribution without employee review.
Use AI as a starting point—not a final judgment
Only 33.5% of Work Contributions included Leadership Principle tags. Employees had to interpret their own evidence and manually map it to an organizational framework, a step that was easy to skip.
01
Generate from grounded evidence
AI analyzes the contribution description and available work artifacts, then recommends up to three relevant Leadership Principles.
02
Keep the employee in control
Employees can accept, reject, edit, regenerate, or manually add principles. Only confirmed selections become part of the official contribution record.
03
Design the failure states
Loading, insufficient-content, retry, and manual fallback states ensure the AI never blocks an employee from completing the contribution.
Product media placeholder
Generate Leadership Principles
Replace with Generate, loading, suggestion, edit, accepted, and error states.
Stable recommendations build trust
During design review, I identified a trust issue: repeated generation could return different LP recommendations even when the employee had changed nothing.
To users, that inconsistency looked like a defect rather than expected model behavior. I advocated for a simple product principle: if the evidence has not changed, the recommendation should not change.
76-86%
Same-result rate for shorter contributions
Repeated tests on identical short inputs were relatively stable, but still not deterministic.
24-31%
Same-result rate for longer contributions
Longer, tag-rich contributions produced much less stable tag sets, creating a high likelihood of visibly different results.
Principle
Regenerate only after meaningful change
I pushed the team to prevent unnecessary regeneration and reserve it for cases where the underlying title, description, or evidence had changed.
Turn collaborators into a contextual feedback loop
The original collaborator field captured who participated in the work, but not the evidence those relationships could provide. I designed the requestor and provider experiences to collect feedback while the work was still fresh.
01
Prompt at a natural moment
When employees add collaborators to a contribution, they can decide whether to request feedback from the people with direct knowledge of the work.
02
Give providers the full context
The request includes the contribution title, description, attachments, and tags so providers can respond to concrete evidence instead of a generic memory prompt.
03
Structure feedback for reuse
I designed questions around how Leadership Principles were demonstrated and what the employee could do next, producing both strength and growth evidence.
Prototype media placeholder
Collaborators + Feedback
Replace with collaborator selection, request feedback, provider response, and manager consumption flows.
Collect once, reuse across the talent journey
Anchoring feedback to the contribution changes it from an isolated annual-review task into evidence that can travel with the work. It also reduces the effort of asking someone to reconstruct context months later.
Employee
Own a stronger career record
Employees retain a project-based portfolio supported by specific observations from collaborators.
Provider
Respond with context in front of them
Collaborators receive a concrete prompt tied to work they directly observed, reducing blank-page effort.
Manager
Review evidence without rebuilding the story
Feedback arrives pre-contextualized and connected to the contribution used during performance decisions.
Early usage showed that a focused warm start could change creation behavior
The Suggested Work Contributions pilot produced a clear early engagement signal. Because impressions may include repeat views, I report actions directly rather than presenting them as a user conversion rate.
7,576
Suggestion impressions
Employees encountered pre-drafted contribution suggestions during the observed pilot period.
998
Create-draft actions
Employees acted on suggestions to create Work Contribution drafts during the same period.
148
Record single-day create actions
Shortly after launch, Warm Start surpassed the previous daily high of 142 create actions before the workday had ended.
Impact visualization placeholder
Warm Start pilot telemetry
Replace with an anonymized chart showing impressions, create-draft actions, and the record usage day.
Small features can create leverage when they connect a larger system
This project reinforced that useful AI does not need to automate an entire workflow. It can remove the hardest first step while preserving the employee's ownership of the final narrative.
01
AI trust is interaction design
Predictability, user confirmation, and understandable failure states mattered as much as the quality of the recommendation itself.
02
Evidence becomes more valuable when it compounds
Connecting capture, articulation, and feedback turns a static project record into an asset that can support multiple future decisions.
03
Measure each layer honestly
The next step is to connect suggestion engagement to draft completion, LP acceptance, feedback response, and downstream manager usage.