Make numbers talk, then catch them lying. Thirteen short steps, with two downloads.
Leading a team? We run this as live training, with your people practicing on their own work.
Ask about team training →Whole quest: 60–90 minutes
There are three kinds of lies: lies, damned lies, and the chart someone made in thirty seconds. This quest you question your data like a lawyer who suspects the witness was coached.
By the end you'll be able to turn a spreadsheet into a chart you can click and filter by describing it, and you'll have a nose for the chart that looks impressive but tells the wrong story.
A CSV is a plain spreadsheet file, nothing fancy. Grab both now.
Before you start, in Claude
Check Settings → Capabilities and make sure Code execution and file creation is on. It usually is. That's what lets Claude analyze a file and build charts you can click.
Before you start, in ChatGPT
The free plan limits uploads to a few files a day, so upload both CSVs today rather than one now and one tomorrow.
Before you start, in Gemini
Nothing to set up. You'll upload files with the + (add files) button in the chat box.
Before you start, in Grok
Nothing to set up. Grok accepts CSV uploads in the chat box.
Before you start, in Muse
Nothing to set up. Muse reads CSV files you attach in chat.
Before you start
Check that your tool accepts file uploads in chat (look for a paperclip or + button). If it doesn't, try Gemini with the Google account from Quest 1.
Pick your AI from the bar at the top of the page and this step will show the exact clicks.
About 3 minutes
About this dataset: Boston neighborhoods, an estimate of their Dunkin' counts, and how many notable Revolutionary-era figures are buried there. The revolutionaries are real and sourced (names and burying grounds are in the file). The Dunkin' numbers are eyeballed, because nobody publishes them by neighborhood. Keep that difference in your back pocket.
Start a new chat in your AI, upload boston_dunkin_revolutionaries.csv, and send:
What's in this dataset? Give me the one-sentence version.
About 5 minutes
Make a bar chart of Dunkin' shops by neighborhood.
Look at it. Then ask the follow-up, which matters more than the chart:
What story is this chart telling? What is it leaving out?
If you get a table or a block of code instead of a picture, reply with "show it as a chart." Some tools need the nudge.
About 5 minutes
Now plot dead revolutionaries against Dunkin' shops. Is there a relationship?
Read its answer, then ask yourself out loud: do I believe that? What's going on here?
Hint: think about which neighborhoods existed in 1776, and which ones are mostly landfill from the 1800s.
About 5 minutes
Show me that same data as a pie chart.
What does the pie make easy to see, and what does it make impossible?
What's a better way to show this data than anything I've tried so far, and why?
Let it make its case. You don't have to agree.
About 10 minutes
Now the thing you couldn't do this morning:
Build me an interactive version I can explore. Let me hover over each neighborhood and filter the data.
In Claude
It should build the chart right in the conversation, ready to click.
In ChatGPT
If you get a static image, reply: "Give me this as a single HTML file I can download and open in my browser." Then open the file.
In Gemini
If it gives you a static chart, ask it to build the interactive version in Canvas.
In Grok
If you get a static chart, ask Grok to build it as an interactive dashboard (Grok Build makes these on every plan).
In Muse
Muse can build interactive pages you open in your browser. If it hands you a static chart, ask for the interactive version.
In your AI
If you get a static image, ask for "a single HTML file I can download and open in my browser," then open it. (HTML is the kind of file a web page is made of.)
Pick your AI from the bar at the top of the page and this step will show the exact clicks.
One last look. You charted a column sourced down to the names on the headstones next to a column of guesses. Did any chart tell you which was which?
About 5 minutes
Pemberton Pickle Works is a made-up specialty pickle company with about 30 people in five departments: Production, Sales & Wholesale, Marketing, Operations & Shipping, and Leadership. They ran a staff survey about AI twice, once before any training (the "pre" wave) and once after it (the "post" wave). Every row is invented. Your job is to turn 59 rows of messy answers into something that tells Pemberton's owners the truth.
Start a new chat, upload pemberton_pickle_pulse_data.csv, and ask for the one-sentence version again to get your bearings.
About 10 minutes
The seven feelings questions use a five-point agree/disagree scale. Get your AI to show you how the answers are spread out, and don't settle for averages.
Hint: try splitting the answers by department, and compare before with after.
About 5 minutes
Before versus after. Did anything move? Ask for the chart that makes change obvious at a glance, then ask whether a different chart would tell a different story.
About 10 minutes
Thirty people also wrote hopes and worries in their own words. Have your AI group the comments into themes. Then ask it for the one quote that would land hardest in front of Pemberton's owners.
Before you trust its themes, read twenty or so of the raw comments yourself. Did it miss a theme you can see?
10–15 minutes
Put together one thing you'd sign your name to in front of Pemberton's owners. An interactive dashboard works, and so does a one-page findings report.
Then the boss move. Keep this one out of the chat. By now you've made a pile of charts. Before you'd show any of them to leadership, decide which one tells the true story and which one only looks impressive. Which would mislead them without anyone noticing? Pick what you'd present and be ready to say why. Bring that to your reflection.
About 5 minutes
Open your recorder and answer the three questions out loud:
Upload it to Bootcamp Reflections and title it "Quest 4 Reflection." That's the last one before the finale.
You turned a messy survey into something that tells the truth. Pemberton was the dress rehearsal for your next real one.
If this felt like Quest 3 again, it was. A chart can flatter you as easily as a sentence can, and now you know how to check which kind you're looking at.
Gemini Notebook, from Quest 1, can search the web for sources on a topic you describe, and on accounts that have it, turn what it finds into a Data Table you can export to Google Sheets. That means you can send AI to find a dataset, then chart it with everything from today. Good targets are the public data people in your field argue about, or your own (your music history, your step count, a quick Google Form sent to friends).
Bring the Quest 3 reflex along. A table AI builds for you is a first draft. Spot-check it against the original sources before you chart it, and look for the places it rounded, guessed, or dropped a row.
If you go down this hole, record a second reflection. Then go touch grass somewhere without Wi-Fi.
The Exit Interview is where you explain what you learned to a tough panel and leave with a Field Report. It's the shortest part of the course and the one people remember.
Bootcamp is the do-it-yourself version of what we run for organizations. In the live version your people practice these habits on their own work, with us in the room. And if your team is anything like Pemberton's, a survey before and after is where you find out what they think.
Ask about team training →Tell us what's going on with AI where you work and we'll point you at the right first step, even if it isn't one of ours.
Subscribe to our biweekly newsletter
Plain-English thinking on AI and people. No spam, easy exit.
You're at Quest 4 · 0% of the way through
It knows where you are on the page and what's nearby. The people behind it are one click away.