Personalized practice, powered by agents

Context

After strong engagement during live sessions, many Varsity Tutors students lost momentum between meetings. Parents wanted evidence of progress, but students were met with generic homework that didn’t reflect what they’d just learned.

I explored how we could use Anthropic’s Workbench and Claude to generate session-specific practice problems and autonomously build matching UI. This meant redefining what “post-session UX” could be—moving from static homework to an intelligent, agent-driven experience personalized by age, subject, and learning style.

Goals

• Increase % of students who complete practice after a session

• Lift 37-day retention and “second-session-within-14-days” rates

• Test agentic personalization for scalability and educational integrity

Process

I began by mapping how session summaries, student metadata, and concept tagging could form the contextual foundation for AI-driven practice. The goal was to make practice problems feel like a natural continuation of the tutoring experience—not a disconnected task.

Working closely with engineering, I designed a three-prompt pipeline that powered each personalized experience:

• Session Analysis Prompt – Extracts learning insights from the transcript

• Content Generation Prompt – Creates targeted problems or games

• UI Generation Prompt – Produces a React component using Varsity Tutors’ design tokens

This “analysis → content → UI” chain ensured every experience was personalized yet visually consistent.

Using Anthropic Workbench, I created structured prompts that turned student data into both learning parameters and design constraints.

To balance AI autonomy with UX quality, I led the creation of design-token guardrails that constrained color, spacing, and typography. We explored three approaches for how much design control to give the agent:

• Branded Templates — Pre-built layouts for each learning modality

  • Trade-off: Stable and fast to implement, but can feel repetitive

• Design System + AI Layouts — Agent composes layouts within established design-token rules

  • Trade-off: Balanced personalization while maintaining visual and brand consistency

• Full AI Generation — Agent freely generates layout and style

  • Trade-off: Highly engaging, but risks inconsistency and accessibility issues

We launched with Option 2, allowing personalized presentation while preserving brand integrity and accessibility standards.

My Role

• Defined the agentic UX strategy and the cross-functional roadmap

• Partnered with engineers to build the prompt architecture and Workbench flow

• Created the UI library and token rules that guided AI-generated layouts

• Facilitated design reviews and prompt experiments with PMs

• Presented the end-to-end prototype to leadership using v0

This project combined hands-on design with organizational alignment—illustrating how AI can extend, not replace, the designer’s role.

Outcomes

• Delivered the first working prototype linking tutoring data → AI-generated content → usable UI

• Established the prompt-chain and design-system constraint model