# Starter prompts Copy-paste prompts for the hands-on exercise. Use them as written, or steal from them. Each prompt is designed to drop into Cursor agent mode or Claude.ai with minimal editing. --- ## Prompt 0. Critique my three exec questions **Use in**: Claude.ai or Cursor (Ask mode) **When**: you drafted your three questions in `src/App.tsx` and want pressure-testing before you build charts against them ``` I'm playing the role of a User Researcher at Starbucks corporate. I have to write three questions that a Starbucks VP would actually pay me to answer using 2,000 Yelp reviews. Here are my drafts: 1. 2. 3. Critique each one as if you were the VP. For each: - Is the question specific enough that a chart could answer it? Or is it too vague to know when I've answered it? - Is it about a real Starbucks decision someone could act on this quarter, or is it a curiosity? - If the question is weak, rewrite it to be stronger. Keep the same topic. End with: which of the three questions is the strongest, and why. ``` --- ## Prompt 1a. Find candidate themes the starter codebook misses **Use in**: Cursor agent mode in the project (it needs file access) **When**: Task 2, step 1 — before you draft a new theme, identify a real gap in the existing codebook ``` Open `public/starbucks_reviews_coded.csv`. Filter to 1-star and 2-star reviews. Read the `text` column for 30 of them, picked at random across cities. The existing codebook has these 10 themes: mobile_app_order, speed_wait_time, drink_quality, barista_service, atmosphere_workspace, drive_thru, cleanliness, food_quality, pricing_value, rewards_loyalty. Tell me the top 3 recurring complaints from these 30 reviews that are NOT already covered by those 10 themes. For each candidate: - A short name for the theme (5 words or fewer). - A one-sentence description of what it captures. - 3 verbatim example phrases pulled from the reviews you read. Be conservative. If a complaint is already covered by an existing theme, do not list it as a candidate — say so and move on. ``` --- ## Prompt 1. Extend the starter codebook **Use in**: Claude.ai or Cursor (Ask mode) **When**: after you've read 20-30 reviews from your chosen city and spotted a recurring topic the starter codebook misses ``` You are a senior user researcher. Below is our starter codebook for analyzing Starbucks Yelp reviews: And here are 20 reviews I just read that mention something the codebook does not currently capture: Propose ONE new theme to add to the codebook. Return a JSON object with these fields, matching the existing schema: - id: snake_case - name: display name - definition: 1-2 sentences. Distinguish it from existing themes. - examples: 3 verbatim phrases from the reviews above that would trigger this theme - business_question: which corporate decision-maker cares about this theme and what would they want to know Be conservative. If the topic is fully covered by an existing theme, say so and stop. ``` --- ## Prompt 2. Re-run thematic coding on a small subset **Use in**: Cursor agent mode in the project (so it can read files and run Python) **When**: after you extended the codebook, to see how the addition changes the picture ``` Take the first 100 reviews from data/starbucks_reviews.csv. Apply my updated codebook at data/starbucks_codebook_starter.json. For each review: 1. Read the review text. 2. Decide which themes apply (zero, one, or several). 3. For each applied theme, find the exact verbatim quote span that supports it (max ~25 words). If you can't find a verbatim quote, do not apply the theme. 4. Output a row with: review_id, stars, city, themes (pipe-separated theme IDs), themes_with_quotes_json. Write the output to data/starbucks_reviews_coded_100.csv. Show me the count of reviews per theme at the end, and flag any reviews where you considered applying my new theme but ultimately did not. ``` --- ## Prompt 3. Generate a chart **Use in**: Cursor agent mode, inside example-dashboard/ or your fork of the starter template **When**: when you have a specific exec question you want to chart ``` I want to add a chart to my dashboard that answers this exec question: "" The data is at ../data/starbucks_reviews_coded.csv. The schema is: review_id, business_id, business_name, city, state, stars, date, text, themes (pipe-separated), themes_with_quotes_json. Constraints: - React + recharts (matching the existing dashboard). - The chart should be filterable by the same filters as the other charts on this page (city, star rating, theme). - Include a 1-sentence chart title that names the exec question. - Show one representative quote beneath the chart when the user clicks a bar. Generate the chart component and wire it into the existing page layout. ``` --- ## Prompt 4. Ground an insight bullet **Use in**: Claude.ai **When**: writing the insight memo and you want to pressure-test that your claim is grounded ``` I'm writing a 1-page research memo for Starbucks corporate. Here is one of my insight bullets: "" Here is the supporting data: Critique this bullet as if you were the Starbucks VP of Research: 1. Is the claim defensible from this data, or am I overreaching? 2. Are the supporting quotes actually about what I claim they are about? 3. What's the strongest counter-question a skeptical exec would ask? 4. Rewrite the bullet in a more defensible way if needed. ``` --- ## Prompt 5. Live exec Q&A — ad-hoc analysis **Use in**: Cursor agent mode during the 4:10-4:25 Q&A block **When**: the VP asks you a question your dashboard doesn't directly answer ``` The Starbucks VP just asked me: "" The data is at ../data/starbucks_reviews_coded.csv. Schema reminder: review_id, business_id, business_name, city, state, stars, date, text, themes. Help me answer in under 60 seconds. Do the analysis, return a single number or short answer, and pull 2 representative quotes that ground it. Keep the analysis simple — I need to be able to defend the method live. ``` --- ## Prompt 6. Stress-test your dashboard before the Q&A **Use in**: Claude.ai with the link to your dashboard repo, or paste your code **When**: 10 minutes before the Q&A starts, if you want a pre-flight check ``` Below is the code for my Starbucks reviews dashboard: Imagine you are the Starbucks VP of Research about to interview the person who built this. List the 5 toughest questions you would ask them. For each question, predict whether the dashboard as-is can answer it directly, and if not, what the researcher would need to do (Cursor agent? new filter? new chart?) to answer in under a minute. ``` --- ## Tips - **Run prompts in order.** Each one builds on the previous. Skipping prompt 1 means your dashboard answers the same questions everyone else's does. - **Iterate inside each prompt.** Generated output is a draft. Spot something off? Reply "actually, also check X" — don't start over. - **Save what works.** When the AI gives you a good chart or a good critique, copy the prompt + output into a `prompts-i-kept.md` file. This becomes your prompt library for the next portfolio piece.