Skip to content
May 22, 2026 · Taste Test

From Reviews to Insights: AI-Assisted UX Research

Take 2,000 real Starbucks reviews from the Yelp Open Dataset and produce a portfolio-grade artifact: an interactive insight dashboard with quote-grounded findings, plus a one-page memo. Then defend your work live in an executive Q&A.

Presenter: Zach Schendel (Head of Research, Google DeepMind) · Lucas 306 · 1:30 to 5:00 p.m. · Tutorial at 2:30

Watch the recording

Watch on YouTube →

Learning goals

By the end of the session, you should be able to:

  • Frame any consumer dataset around an executive's business question first, not the data you happen to have.
  • Build a thematic codebook with definitions, examples, and the business question each theme answers — and use AI to apply it without hallucinating.
  • Ground every insight in a verbatim customer quote so a skeptical exec can trace it back to source.
  • Vibe-code an interactive dashboard with Cursor that you can re-query live to answer questions you didn't anticipate.
  • Apply the four recipes (business question, codebook, citations, dashboard) to any qual+quant dataset you encounter next.

Starter kit

Grab these before the exercise. The starter template runs on localhost out of the box; you'll extend it with Cursor Pro during the hands-on block.

.zip · Starter template
Vite + React + recharts dashboard
One working chart, the rest is TODOs. Runs on localhost with npm install && npm run dev. You build the rest.
.csv · Data
2,000 coded Starbucks reviews
Stratified sample from the Yelp Open Dataset. Already thematically coded against the starter codebook with verbatim quotes.
.json · Codebook
10-theme starter codebook
Each theme has an id, definition, examples, and the business question it answers. Extend it during the exercise.
.zip · Reference
Example dashboard (finished)
What "done" looks like: four charts, filter panel, quote list, Cursor agent hint. Use as a target, not a copy.
.pdf · Handout
Four recipes one-pager
The transferable principles. Print it and stick it above your desk for your next project.
.md · Template
One-page insight memo
Three findings, three quotes, three "so whats". The format every PM and VP actually reads.
.md · Prompts
Copy-paste starter prompts
Eight prompts: exec-question critique, codebook gap finder, codebook extension, chart generation, insight grounding, stress-test, and live exec Q&A queries — pre-written for Cursor and Claude.
.md · Q&A bank
15 plausible exec questions
The kind a Starbucks VP would actually ask. Practice with these before the live Q&A rehearsal at 4:10.

Tutorial outline

Two hours, six blocks, one artifact. The session moves quickly: ~30 min of tutorial, 50 min of hands-on, 15 min of live Q&A rehearsal, and a portfolio-takeaway wrap.

2:30-2:35
Welcome + frame
Why "AI fluency" in research means more than running a chatbot. The artifact you're walking out with.
2:35-2:50
Industry preview
My team's real version — what the workflow looks like at scale on real product data — to set the ceiling.
2:50-3:20
Live tutorial: Starbucks end-to-end
Business question → codebook → thematic coding with verbatim citations → dashboard → quote-grounded insight. The four recipes surface as we go.
3:20-4:10
Hands-on build (50 min)
Four tasks in the starter template: write exec questions, extend the codebook, add two charts, build the quote panel and draft a memo.
4:10-4:25
Live exec Q&A rehearsal
Three volunteers defend their dashboards live. Audience plays VP of Customer Experience, Director of Stores, and a skeptical CFO.
4:25-4:30
Wrap + portfolio takeaway
How to put this in a portfolio (and what hiring managers will look for). One thing to try this week.

The four recipes

Transferable techniques, tool-agnostic. They work whether you're using Cursor, Claude, ChatGPT, or whatever ships next quarter. Each one shows up explicitly in the dashboard and the exercise.

Recipe 1
Start from the business question
Not the data. "What's hurting drive-thru speed?" is a research project. A spreadsheet is a spreadsheet.
Recipe 2
Codebook, not vibes
Themes get definitions, examples, and the question they answer — before you let AI touch the data. This is what makes qual countable.
Recipe 3
Cite or it didn't happen
Every theme assignment carries a verbatim quote. A local verifier drops any AI output where the quote isn't in the source. No anecdotes, no hallucinations.
Recipe 4
The dashboard is a question machine
Don't ship a static report. Ship a tool where an exec can ask a new question and you have an answer in 60 seconds — with the supporting quote.

Hands-on exercise

Four tasks in 50 minutes inside the starter template. You'll need Cursor Pro and Claude Pro open. All prompts you'll need are pre-written in materials/starter-prompts.md.

Task 1 · Write three exec questions

Target time: 5 min · Recipe 1 in the wild
  1. Pick a Starbucks exec persona: VP of Customer Experience, Director of Stores Operations, or CFO.
  2. Write three questions that exec would actually ask before a quarterly review. Be specific. "How are we doing?" doesn't count.
  3. Drop them into App.tsx → QuestionStrip. Every chart you add later has to answer one of these.
  4. Use the "Critique my three exec questions" prompt (Prompt 0 in starter-prompts.md) to pressure-test your questions before you build charts against them.

Task 2 · Extend the codebook

Target time: 10 min · Recipe 2 in the wild
  1. Open the codebook (public/starbucks_codebook.json) and scan the 10 starter themes.
  2. Use the "Find candidate themes" prompt (Prompt 1a in starter-prompts.md) to have Cursor scan the coded CSV and surface gaps in the starter codebook.
  3. Open Claude or Cursor and paste the "Extend the starter codebook" prompt (Prompt 1) from starter-prompts.md. Drop in ~20 reviews as raw context.
  4. Add 2-3 new themes that are clearly present in the data but missing from the starter. Each new theme needs an id, name, definition, two examples, and the business question it answers.
  5. (Stretch) Re-code a subset of reviews against your extended codebook using scripts/02_run_thematic_coding.py.
Resist the urge to invent themes that "should" be there. Only add a theme if you can find at least three reviews that clearly support it. Conservative coding builds executive trust.

Task 3 · Add two more charts

Target time: 25 min · Recipe 4 in the wild
  1. The starter template has one working chart (theme frequency) and three TODO slots.
  2. Pick two of: stars-by-theme stacked bar, theme-by-city grouped bar, top complaints in 1-2 star reviews, or invent your own.
  3. Use Cursor agent mode. Paste the "add a chart" prompt from starter-prompts.md and let Cursor scaffold the component. Then iterate.
  4. Each chart must answer one of the exec questions you wrote in Task 1. Label it explicitly on the chart card.

Task 4 · Quote panel + memo

Target time: 10 min · Recipe 3 in the wild
  1. Wire up the quote panel: clicking a theme bar (or a row in your stacked chart) should filter the quote list to verbatim spans for that theme.
  2. Open materials/insight-memo-template.md and fill in the three findings. Each finding gets a "what we see" sentence, the supporting quote, and a "so what" recommendation.
  3. Save the memo next to your dashboard URL. Together they are the portfolio artifact.
Three volunteers will defend their dashboards live in the 4:10 exec Q&A rehearsal. Zach plays three exec personas: VP of Customer Experience, VP of Operations, VP of Research. The questions are in exec-questions-bank.md — but the questions you'll actually get are the ones you didn't plan for. That's the test.

Going further

Session led by Zach Schendel, Head of Research at Google DeepMind and adjunct faculty at Santa Clara University.
Host: Kai Lukoff · klukoff@scu.edu.