4. AI Tutors
Rice Business 2026
Rice Business
Our AI mission: To explore strange new worlds, to seek out new life and new civilizations, to boldly go where no one has gone before
Four Years, Chatbot to Agent
AI has been truly useful for less than 18 months
Andrej Karpathy, OpenAI co-founder, former Tesla AI chief, now at Anthropic
October 2025
“They just don’t work. They don’t have enough intelligence, they’re not multimodal enough, they can’t do computer use and all this stuff. They don’t have continual learning. You can’t just tell them something and they’ll remember it.”
“I feel like the industry is making too big of a jump and is trying to pretend like this is amazing, and it’s not. It’s slop.”
December 2025
“I really am mostly programming in English now, a bit sheepishly telling the LLM what code to write… in words. Biggest change to my basic coding workflow in ~2 decades.”
February 2026
“It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the ‘progress as usual’ way, but specifically this last December. Coding agents basically didn’t work before December.”
“I don’t think I’ve typed like a line of code probably since December.”
What HBS Is Doing
Course tutorbots
Live AI case exercises
In-class AI simulations
No-code data analysis
AI feedback on student work
Classroom avatars
Details and sources: rice-workshop.kerryback.com/hbs.html
Today’s Session
1
Data Analysis
2
Tutors
3
Skill Demos
Gemini Notebook (previously NotebookLM)
Free knowledge repository. Upload as many as 50 sources into each notebook. Get a chatbot that answers from sources, narrated slides, podcast-style audio, quizzes, …
- Chatbot example: Options textbook
- Narrated slides example: Kapadia-Ostdiek-Weston paper
- Podcast example: Put-call parity from textbook
Chatbot is great. Videos are good for a short change of pace. We can specify only use this part of the source for video or podcast. Cannot revise directly but can regenerate with new instructions.
Data Analysis
1. Run a Regression
Prompt
Read data.xlsx. Regress Y on X. Show the usual statistics and a scatter plot with the regression line. Save the plot and open it.
2. Create an App
Prompt
Create a single-page app — one self-contained HTML file that runs in the browser, no server and nothing to install — for a user to upload an Excel file containing X and Y columns. Regress Y on X. Show the usual statistics and a scatter plot with the regression line.
3. Refine the App
Prompt
Modify the app so that, if the Excel file has a date or month column, the date appears in the hover data on the scatter plot.
4. Deploy the App
There is a one-time setup needed before deploying (create a free GitHub account, tell Claude to do the rest). The following prompt works for me.
Prompt
Name the app regression.html. Copy it to the docs folder. Commit and push.
Wait a minute or so, then go to rice-workshop.kerryback.com/regression.html
Improve the User Interface
It is probably worthwhile to ask Claude to install the front-end design plugin from Anthropic’s marketplace before building apps.
Prompt
/plugin install frontend-design@claude-plugins-official
Artifacts
This can be done in Claude Desktop (or at claude.ai). Choose Home/Artifacts. Then New Artifact (in Desktop, select either Chat or Cowork).
Prompt
Create an artifact for a user to upload an Excel file containing X and Y columns. Regress Y on X. Show the usual statistics and a scatter plot with the regression line.
Data
Data is generally a problem, but there is quite a bit of free finance data available. This uses my finance-data plugin (/plugin install finance-data@kerryback or Academic Studio/Run Setup))
Prompt
Get the market excess return and risk-free rate from French’s data library for the past 60 months and compute monthly returns for Tesla from Yahoo Finance data. Run the CAPM regression and report the results.
Tutors
Idea
- AI displays page of teaching
- AI asks questions, provides feedback
- Could make stand-alone, per-course or per-topic (need API call for AI)
- Easier to make Claude skill, student’s Claude is AI
Example — Binomial Option Tutor
/plugin marketplace add ricebusiness/plugins/plugin install tutor@ricebusiness- Start a new conversation
- Enter /tutor:binomial
What the Binomial Tutor Does
Four modules
- Replication and no-arbitrage in a one-period tree
- Where the risk-neutral probability comes from
- Two-period and n-period backward induction
- American options and early exercise
Coaching, not answers
- The student produces the numbers; Claude asks and then waits
- A wrong answer is diagnosed from the number itself, and answered with a question
- The whole arc runs 30–45 minutes, and the student can stop between modules
The SKILL.md
What Else Is in the Plugin
Reference files
- One teaching script per module, plus a bank of problems with verified answers
- SKILL.md names them but does not contain them — Claude reads a module only when the student gets there
- Keeps the always-loaded instructions short and the material deep
A Python script
- Prices European and American calls and puts on a tree, and reports the replicating portfolio at every node
- Claude runs it before asserting any number, so it never marks a correct answer wrong
- Can blank out a node to turn any tree into an exercise, and plot the trees
Why it is a plugin, not a prompt
The tutoring protocol, the course material, and the arithmetic live in one installable package that any student can add in two commands, and that you can revise once for everyone.
Example: Agency Tutor
Business law, bundled in the same plugin (buyer beware!)
- Start a new conversation
- Enter /tutor:agency
What the Agency Tutor Does
Four modules
- Actual authority — what the agent was entitled to do
- Apparent authority — what the third party was entitled to believe
- Ratification and estoppel, when there was no authority at all
- Who pays: disclosed, unidentified, and undisclosed principals
Fact patterns, not lectures
- The student gets a one-paragraph problem and has to call it
- A wrong answer is diagnosed from the theory the student named, not corrected
- Runs 40–50 minutes, and stops cleanly between modules
The Same Concept, Different Details
| Binomial tutor | Agency tutor | |
|---|---|---|
| The student produces | A number | A theory, and a call |
| How Claude knows the answer | binomial.py prices the tree |
agency.py builds the fact pattern and the answer key from the same settings |
| How a wrong answer is diagnosed | The number names the mistake | The theory named reveals the step skipped |
| The failure to guard against | Marking a correct answer wrong | Inventing a case citation |
Teaching Tool Demos
Installing the Plugins
Four plugins, one marketplace. Type each line at the Claude Code prompt.
Step 1 — add the marketplace
/plugin marketplace add kerryback/skills
Step 2 — install the four plugins
/plugin install voiceover@kerryback
/plugin install participation@kerryback
/plugin install poll@kerryback
/plugin install screenshare@kerryback
Restart Claude after installing. Each plugin builds its own small Python environment the first time you run it, so the first launch is slower than the ones after.
Demo 1 — Narrated Video from a Deck
What it does
- You give it the PDF of your slides — nothing else
- Claude writes the narration; you edit any line you don’t like
- ElevenLabs voices it into an
.mp4plus a.txttranscript
What you do
- Download slide_rule.pdf
- Enter /voiceover slide_rule.pdf
- Read the narration, rewrite one slide, then generate
Needs a free ElevenLabs API key, pasted into the banner at the top of the app. Seven slides fits inside the free monthly quota. Your students receive an mp4 to watch.
Demo 2 — Participation from a Photo Roster
What it does
- Imports an ESTHER course roster PDF — every name and headshot
- Tap 1–3 for amount and 1–3 for quality, add a note, set the date
- Save writes one row per student to
participation.csvin the course folder
What you do
- Save the roster PDF from the Zoom chat, or pull your own from ESTHER
- Enter /participation
- Set up a class, import the roster, score three students, open the CSV
The sample roster is a real one, so it goes out in the chat rather than on the course site. Claude’s only job is to start the app — it launches a local web server on your machine and tells it where the PDF is. The app reads the roster itself, so the names and headshots never pass through a model and never leave your laptop.
Demo 3 — A Live Poll on Their Phones
What it does
- Type the question as a sentence; Claude works out what kind of poll it is
- Multiple choice, word cloud, confidence scale, numeric estimate, ranking
- Anonymous by design — no names, no roster, nothing stored per student
What you do
- At start of class enter /poll in Claude in the classroom computer. Brings up a QR code in browser.
- Students connect
- During class, either /poll
<filename>for prepared poll or /poll<question>for on-the-spot - Needs
cloudflaredon classroom computer, and I have to check it is not blocked (workaround exists)
/poll How confident are you using AI in your teaching? 1 to 5
Demo 4 — Student Screen Shares
- Enter /screenshare on the classroom computer
- Show the browser page that opens — join link and room code on the page
- Students connect on their laptops
- Return to the page and select a student to share whenever you want
Also needs cloudflared, but workaround exists if blocked