How We Can Use AI in Teaching and Research

Lunch and Learn — Session 2

Kerry Back, Rice University

The AI Agent Landscape

OpenAI Codex

• Terminal-based coding agent
• Full local file access
• Runs code, edits files, uses tools
• Free with OpenAI account

Google Gemini CLI

• Terminal-based agent
• Google ecosystem integration
• Code execution, file access
• Free tier available

Anthropic Claude Code

• Terminal-based universal agent
• Full local file access + internet
• Skills, MCP connectors, subagents
• Also available as Claude Desktop (Code tab)

All three are AI agents with tools — not chatbots. This session focuses on Claude Code and Claude Desktop, but the concepts (skills, connectors, iteration) transfer to any agent platform.

I want to start by acknowledging the landscape. OpenAI, Google, and Anthropic all offer AI agents with tools — not just chatbots. I’ll focus on Claude Code and Claude Desktop because that’s what I use, but the concepts transfer. The key point: these are all agents, not chatbots.

Claude Desktop: Three Modes

Chat (Analysis)

• Sandboxed Python in the browser
• No local file access
• Must upload data manually
• Good for quick analysis

Cowork

• Runs in a local VM
• Files synced to/from VM
• No internet access (sandbox)
• Safe for sensitive work

Code

• Runs on your machine
• Full local file access
• Full internet access
• Same engine as Claude Code CLI

The Code tab in Claude Desktop and the terminal command claude use the same engine. Code mode gives Claude direct access to your machine — no sandbox, no VM, no upload step.

Claude Desktop has three modes. Chat is sandboxed — good for quick questions. Cowork runs in a VM with no internet — safe for sensitive data. Code runs directly on your machine with full access. The Code tab and the terminal command claude are the same engine.

Claude Code: Universal Agent

What makes it universal

• Full access to your local file system, internet, and installed tools
• Add instructions via CLAUDE.md — Claude reads this on every startup
• Add capabilities via skills — text files Claude reads only when needed
• Add external tools via MCP connectors — email, calendar, databases, browsers
• Run subagents in parallel for complex tasks
• One interface for everything

Claude Code is a universal agent. You extend it with text files (CLAUDE.md, skills) and connectors (MCP). The result is a single conversational interface for your entire workflow — slides, email, grading, research, file management.

General Tools

Let’s start with everyday tools — email, calendar, task management, and file search, all through a single conversational interface.

Email & Calendar

Gmail MCP Connector

• Reads both email accounts (Rice institutional + personal)
• Manages both calendars — creates events, checks conflicts
• Drafts replies, identifies emails needing responses
• Claude walks you through the one-time setup

“Are there any events in my emails that should be added to my calendars? If so, add them.”

“Are there any emails that I need to reply to?”

“Draft an email to [person] about [topic] saying [key points].”

Rice’s security policy makes merging email accounts difficult, but it doesn’t restrict API access. Claude can read and manage both accounts from a single conversational interface.

I have two Gmail accounts — Rice and personal — and Rice’s security policy makes merging them difficult. But the security policy doesn’t restrict API access. Claude reads both, manages both calendars, drafts replies. One interface instead of switching between accounts.

Task Management

One fewer app to manage

• Ask Claude to create a task skill with categories: personal, teaching, research
• “What are my open teaching tasks?”
• “Add a task to review the midterm exam by Friday”
• “What’s overdue?”
• Stored as files — Claude reads and updates them via the skill

There are many to-do apps, but it simplifies things to do everything inside Claude Code. The way all of personal computing will soon work, I think.

You can keep your task list inside Claude Code. Create a task skill with categories, then manage everything from conversation. One fewer app to switch between.

Git & File Search

Version Control

• Ask Claude to initialize a git repo and commit periodically
• Each commit is a snapshot of your work
• Search past versions: “I used to have a paragraph that discussed … Find it”
• Much easier than searching old Dropbox or Overleaf versions
• Push to GitHub for backup, sharing, or websites (GitHub Pages)

File Search

• “I have a docx somewhere in my Dropbox folder that discusses … Find it”
• Claude searches your entire machine with vague clues
• Works across all file types
• Git history search: find deleted paragraphs, old versions of analyses
• You don’t need to know git — Claude handles the commands

Git is great but complex. Fortunately, Claude knows all about it, so you don’t need to know anything. Ask Claude to initialize a repo, commit periodically, and search past versions. File search across your entire machine with vague clues is also extremely useful.

AI for Research

Now let me shift to the research side — from literature review through data analysis to writing and revision.

Literature Review & Search

AI-Powered Search

• Claude Code searches the web, downloads papers, summarizes findings
• Spawn parallel subagents: “Search for papers on these three topics simultaneously”
• AI reads and summarizes papers, identifies methodology gaps
• Compare year-over-year changes in 10-K disclosures or policy documents

NotebookLM

• Upload up to 50 sources — papers, reports, transcripts
• Query across all sources with inline citations
• Audio Overview generates a podcast-style discussion of your research
• Great for synthesizing a literature before writing the introduction

Two complementary approaches. Claude Code does active search — finds papers, downloads them, summarizes. NotebookLM is for deep synthesis — upload your 50 key papers and query across all of them with citations. Use both.

Planning & Model Building

Plan Mode

• Shift+Tab to force Claude into plan mode
• Claude also enters plan mode automatically for complex projects
• In plan mode, Claude researches and architects — editing is disabled
• Claude asks clarifying questions before execution
• Plans are saved to ~/.claude/plans/ — persisted on disk

Model Building

• Describe the model in words — Claude writes the code
• Specify your preferred language (Python, R, Stata, MATLAB)
• Iterative refinement: “add fixed effects,” “cluster standard errors by firm”
• Claude writes reusable scripts — scripts are the artifact, not just the output
• Build replicable pipelines, not ad-hoc commands

Planning is as important as execution. For substantial projects, invest in the plan — then clear context so implementation has a clean slate. The plan file is Claude’s instructions to itself.

Shift+Tab forces plan mode. Claude researches, asks questions, and designs the approach — no code is written yet. After planning, clear context so implementation starts fresh. For model building, describe what you want in words and iterate.

Data Handling & Analysis

Data Acquisition

• Claude finds, downloads, and parses data from APIs, websites, and files
• Handles messy Excel structures, scrapes SEC Edgar, Census, FRED
• Set user-agent headers to avoid HTTP 403 blocks on government sites
• Cache raw data locally — re-runs only download new/missing files

Large Datasets

• Convert CSV → Parquet for dramatic compression (70 GB → 1 GB)
• Query with DuckDB — serverless SQL, no database setup
• Schema harmonization across eras (e.g., HMDA changed column names in 2018)
• Add metadata and data dictionaries to help future sessions
• Include resume capability in download scripts

Claude samples data to understand structure, then writes a script to process the full dataset. It never loads everything into the context window. DuckDB + Parquet is the modern stack for research data — fast, portable, and free.

Claude inspects a small portion of data to understand the structure, then writes a reusable script. For large datasets, convert to Parquet and query with DuckDB — a 70 GB dataset becomes 1 GB. Always cache raw data and build replicable download scripts.

Generating Figures & Tables

From verbal description to publication-ready output

• “Plot homeownership rates by age over time using Kieran Healy’s best practices” — Claude knows published style guides
• Generates ggplot2 (R) or matplotlib/seaborn (Python) figures
• Iterate: “make the legend larger,” “use direct labeling instead of a legend,” “switch to a coefficient plot”
• Formatted regression tables (stargazer, modelsummary, estout)
• Inserts figures and tables directly into your LaTeX manuscript

Start rough, review, refine. Reference style authorities by name — Claude knows Kieran Healy’s Data Visualization, Tufte’s principles, and journal-specific formatting requirements.

You don’t need to specify every detail upfront. Start rough, review, then refine. Claude knows published visualization books — say “follow Kieran Healy’s best practices” and it applies clean ggplot2, direct labeling, minimal chart junk. Tables go straight into LaTeX.

Writing LaTeX

Drafting & Editing

• Claude writes LaTeX sections from verbal descriptions or outlines
• Inserts \includegraphics and table environments referencing your generated outputs
• Compiles the PDF and reviews the result autonomously
• Build a personal style guide skill from your published papers — Claude drafts in your voice

Revision Workflows

• LLM as editor, not writer: “Create inline comments where my argument is poor — do not edit my text”
• Preserves your voice — prevents convergence to generic AI prose
• Strategic revision skill (Jukka Sihvonen): upload manuscript + referee reports → DAG-validated revision plan
• Identifies conflicting referee demands and parallelizable tasks

Writing is thinking — be deliberate about what you offload. Use the LLM as an editor and collaborator, not a ghostwriter. 1977 IBM: “A computer can never be held accountable, therefore a computer must never make a management decision.”

Two modes. For drafting, build a style guide skill from your papers — Claude captures your voice, tone, and verbal habits. For revision, use the inline commenting pattern to get feedback without losing your voice. Sihvonen’s strategic revision skill turns referee reports into a DAG-ordered revision plan.

VS Code for Research

VS Code with file explorer, Python editor, LaTeX editor, LaTeX preview, and Claude Code terminal

This is what the research workflow looks like in practice. VS Code with five panes: file explorer on the left, Python code for analysis, LaTeX manuscript, a live preview of the compiled PDF, and the Claude Code terminal at the bottom. One application orchestrating the entire pipeline.

VS Code + Remote Server (jgsrc1)

Setup

• Install the Remote SSH and Claude Code extensions in VS Code
• SSH into jgsrc1 from within VS Code
• Install Claude Code on the server: curl -fsSL https://claude.ai/install.sh | bash
• Run claude in the VS Code terminal

What This Gets You

• Edit files on the server with AI assistance
• Run computations on jgsrc1’s hardware
• Claude Code operates on the server’s file system directly
• VS Code runs on your laptop; everything else runs remotely
• Same experience as local — just on a more powerful machine

If you do computation-heavy work, run Claude Code on the school server. Install Remote SSH in VS Code, connect to jgsrc1, install Claude Code there. Full AI agent on server hardware while you interact from your laptop.

Best Practices

Let me pull together the key principles for working effectively with AI agents in research.

Plan → Execute → Evaluate

📋 Plan Shift+Tab for plan mode. Research, ask questions, architect before coding.

⚙️ Execute Clear context after planning. Implementation starts fresh with the plan file.

🔍 Evaluate Run the critique skill. Three parallel reviewers catch what you’d miss.

Trust but verify: Know your expected output so you can catch errors — e.g., 30 firms × 4 years = ~120 rows. Resist the “drinking bird” temptation of blindly hitting “yes.” Claude is a remarkable programmer but makes mistakes on edge cases.

The three-phase cycle. Plan with plan mode — research and architect before writing code. Execute with a clean context window. Evaluate with the critique skill. And always verify — know your expected output dimensions.

Working with Subagents

What They Are

• Claude spawns independent agents to work on subtasks
• They run in the background — you continue your conversation
• Claude decides when to use them spontaneously (parallelizable tasks)
• Or you can direct: “Use subagents to research these three topics in parallel”
• Each subagent has its own context — doesn’t clog the main conversation

Gotchas

• Subagents do not inherit all parent instructions
• If you told Claude “stay in this folder,” a subagent may still explore elsewhere
• You must re-state constraints at the subagent level
• Explain why it should stop, not just say no — subagents are persistent
• Denying a tool call alone is not enough; the agent will try alternatives

Subagents are how Claude parallelizes work — critique skill, parallel research, file processing. But they don’t inherit all parent instructions. You may need to re-state constraints and explain why. Denying a tool call isn’t enough; explain the reason.

Context Window & Keyboard Shortcuts

Context Window Management

• Everything consumes the context window — being precise improves performance
• /compact — summarizes conversation to free space
• Start new sessions when focus drifts
• Research → plan → implement workflow: write findings to a file, start fresh, read the file
• Subagents protect the main window

Key Shortcuts

• Escape — interrupt; double-escape — roll back
• Ctrl+O — toggle Claude’s thinking
• Shift+Tab — force plan mode
• @filename — reference a file
• Tab on a permission prompt — add extra instructions
• Up arrow — edit previous message

The context window is like working memory — keeping it focused is the single biggest factor in output quality. Use /compact, start fresh sessions, delegate to subagents. And learn the shortcuts: Escape, Ctrl+O, Shift+Tab, @ for files.

Data Privacy

Treat Claude Code’s access like Dropbox

• Everything in your conversation — prompts, file contents, tool outputs — goes to Anthropic’s API
• Do not expose: PII, HIPAA data, API keys, passwords, IRB-restricted data
• If you accidentally paste a secret, delete it and rotate the credential immediately
• Wall off sensitive directories — don’t launch Claude Code from a folder containing restricted data
• FERPA: anonymize student submissions before AI grading

Claude Code runs on your machine but sends context to Anthropic’s servers. The same caution you’d use with Dropbox or email applies here. Be deliberate about what enters the conversation.

Essential for a university setting. Claude Code runs locally but sends conversation context to Anthropic’s API. Don’t paste API keys, student data, or IRB-restricted information. If you accidentally expose a credential, rotate it immediately.

Sycophancy & the Critique Skill

Reviewer 1: Correctness

• Factual errors?
• Logical gaps?
• Missing information?
• Claims supported?

Reviewer 2: Clarity

• Logical structure?
• Anything confusing or buried?
• Main message direct enough?
• Redundancy?

Reviewer 3: Devil’s Advocate

• Strongest counterarguments?
• Weakest reasoning?
• What would a skeptic challenge?
• Alternative interpretations?

Sycophancy warning: Claude’s feedback on your work will be relentlessly positive — even the weakest argument gets praised. Push hard: “Be harsh. What would a skeptical reviewer attack?” The critique skill helps by assigning an explicit devil’s-advocate role.

Invoke with /critique filename. Three parallel reviewers. The devil’s advocate role is essential because Claude defaults to relentless positivity. You have to explicitly ask for harsh criticism to get useful feedback.

The Autonomous Review Loop

✏️ Create Claude writes the document, code, or slides

🔄 Render Compile, execute, or build the output

👁️ Review Claude inspects the rendered result visually

🔧 Fix & Repeat Edit and re-render until no issues remain

Don’t review AI’s first draft. Let it critique itself and revise autonomously. Repeat until stable. Then review the final version. Your time is expensive — AI’s time is a few cents per cycle.

This loop applies to everything: LaTeX papers, slide decks, code output, figures. Claude creates, renders, inspects, and fixes without prompting. You review a final draft instead of a first draft.

This is the single most important pattern. Don’t review the first draft — that’s the worst version. Let Claude iterate autonomously: create, render, review the output, fix, and repeat. You review the final version. This applies to papers, slides, figures, code — everything.

Teaching with AI

Now let’s turn to teaching — from creating slides to in-class AI use to assessment.

Slide Creation Workflows

Beamer (LaTeX)

• AI writes the .tex file, compiles to PDF
• Review: generates PNG of each slide and inspects for overflow, spacing, alignment
• Edits and repeats autonomously
• Create a skill with your style preferences

PowerPoint

• AI generates .pptx via PptxGenJS (Node.js) and edits via direct XML manipulation
• Review: converts .pptx → PDF via LibreOffice, then PDF → PNGs for visual inspection
• Good for university or department templates
• Can read and modify existing decks

Quarto → reveal.js

• Markdown-based — simpler than LaTeX, more versatile
• Embed executable code — figures generated at render time
• Review: uses the browser MCP tool to view rendered HTML slides directly
• Accepts LaTeX math notation

The autonomous review loop applies to all three formats. The key difference is how Claude sees the output: PNGs for PDF-based formats, the browser tool for HTML. Claude creates → renders → reviews → fixes → repeats until satisfied.

Three paths, three review methods. Beamer: generate PNGs from the PDF. PowerPoint: convert to PDF via LibreOffice, then to PNGs. Quarto: use the browser MCP tool to view the rendered HTML directly. In all three cases, the autonomous create-render-review-fix loop applies.

Quarto: Embedded Code & Interactive Slides

Embedded Code Blocks

• Write Python or R code directly in your .qmd file
• Figures are generated at render time — always up to date
• Change a parameter, re-render, new figure appears
• Students see the figure (or optionally the code too)
• No separate script files to manage

Interactive Plotly Slides

• Plotly produces interactive HTML charts — hover, zoom, pan
• Embed directly in reveal.js slides
• Students explore data live during lecture
• Efficient frontier, yield curves, return distributions — all interactive
• Works in any browser, no install required

Quarto + Plotly turns static lecture slides into interactive data exploration. Claude writes the code blocks, renders the slides, views them in the browser, and iterates. These slides are a Quarto deck that Claude built.

The killer feature of Quarto is embedded code. Write a Python block that generates a Plotly chart, and it renders as an interactive figure in your HTML slides. Students can hover, zoom, and explore. Claude writes the code, renders, reviews in the browser, and iterates.

VS Code for Teaching

VS Code with file explorer, Quarto source, rendered slide preview, and Claude Code terminal

Here’s the teaching workflow in VS Code. File explorer on the left, Quarto source in the center, a live preview of the rendered slides, and Claude Code at the bottom. Edit the .qmd, preview updates in real time, and Claude handles the code blocks and formatting.

CLAUDE.md and Skills

CLAUDE.md

• A text file in your project folder (or global ~/.claude/CLAUDE.md)
• Contains your preferences, project context, and rules
• Claude follows these instructions automatically — no need to repeat them
• Example: “Always use the metropolis Beamer theme,” “Use Python, not R,” “When generating charts, use blue and amber colors”

Skills

• Text files with reusable instructions — Claude reads them only when needed
• Invoke with /skill_name or let Claude decide automatically
• Claude can create skills for you: “Create a skill for building Beamer decks”
• Examples: beamer-create, critique, grading rubrics, style guides

Think of CLAUDE.md as standing instructions for a research assistant. Skills are specialized playbooks for recurring tasks. Both are just text files — nothing to install.

CLAUDE.md is read on every startup — your permanent preferences. Skills are read on demand — specialized playbooks. Both are just text files. Claude can create skills for you — just describe what you want.

In-Class Use of AI

Chat Exercises

• “Chat with Claude/ChatGPT for ten minutes about the efficient frontier. What did you learn that surprised you?”
• Students engage with concepts at their own pace and level
• AI adapts explanations to the student’s questions
• Low stakes, high engagement — works in any discipline
• Debrief as a class: what did the AI get wrong?

NotebookLM for Research Papers

• Upload assigned research papers to NotebookLM
• Audio Overview generates a podcast-style discussion of the paper
• Students listen before class — arrive with context and questions
• Query the paper in class: “What is the identification strategy?” “What are the limitations?”
• Inline citations back to the source text

These approaches work for any course. The chat exercise takes zero preparation — just a topic and ten minutes. NotebookLM Audio Overviews turn dense papers into accessible pre-class listening.

Two low-friction approaches. First, give students ten minutes to chat with an AI about the topic — then debrief. Zero prep required. Second, use NotebookLM Audio Overviews to turn research papers into podcast-style discussions. Students arrive with context and better questions.

Canvas API

Setup

• Request an API key once (Claude tells you how to find the link in Canvas)
• After that: no more SSO login, no more two-factor authentication
• Claude handles all Canvas interactions via the API

What Claude Can Do

• Upload files and assignments
• Download student submissions
• Upload gradebook and comments
• Help with grading — provide a rubric and let Claude assess
• Ask for a summary of why Claude assigned each grade

FERPA compliance: Anonymize submissions before grading. Claude strips student identity, grades against the rubric, then de-anonymizes to upload scores. Test with a dummy gradebook first.

Canvas has an API. You request a key once, and from then on Claude handles all the file management. For grading, anonymize first, grade with AI assistance, then de-anonymize. Always test with a dummy gradebook before running on real grades.

The AI Oral Examiner

📄 Upload Student submits slides (PDF)

🔍 Pre-Analysis AI identifies key claims, gaps, question areas

🎙️ Voice Exam 10–12 adaptive questions via voice AI

📊 Grading Council Three AI models grade independently, then deliberate

Why oral examination?

• “Did the student do the work?” is now the wrong question — AI writes a report or builds a model in 60 seconds
• The right question: can they defend it?
• Can they explain assumptions? Field hard questions? Catch what AI got wrong?
• Practice mode available 24/7 — students improve through repetition

When AI can produce a polished report or slide deck in under a minute, what are we actually testing? The durable skill is verification and defense. The oral examiner tests whether students understand and can defend their work.

The Presentation Examiner

How It Works

• Students receive a magic-link login (no passwords)
• Upload PDF slides
• Session 1: Present to AI listener (ElevenLabs voice agent)
• Session 2: Answer 3 AI-generated questions + follow-ups
• Claude Sonnet analyzes slides and generates questions
• GPT-4 grades on three rubrics with detailed feedback

What It Solves

• Traditional: schedule every student for a live presentation (nightmare)
• This system: anytime, anywhere presentations
• Instant, consistent grading with detailed rubric feedback
• Instructor reviews transcripts and grades later
• Scales from 20 to 2,000 students without hiring graders
• Live now — piloted at Rice

This is our implementation. Students get a magic-link login, upload slides, and present to an AI voice agent. Claude generates questions from the slide content, GPT-4 grades on three rubrics. The instructor reviews transcripts and grades. It scales without hiring graders.

Concerns and Validation

Cost

• API costs: ~$1 per exam at current pricing
• Development and maintenance: ongoing
• Faculty review time for borderline grades
• Total cost of ownership exceeds API cost
• But scales better than human grading

Fairness

• Practice mode available 24/7
• Extended time accommodations built in
• Text-based alternative available
• Accent/fluency bias testing planned
• FERPA review planned

Validation Plan

• Blind AI vs. faculty grading comparison
• Pre-registered reliability threshold (Cohen’s kappa)
• No deployment without pilot evidence
• Student feedback and learning outcomes
• Target: 2027–28, pending results

Transparent about costs: total cost of ownership exceeds the API cost. Bias testing is planned, not an afterthought. Validation is rigorous: blind comparison, pre-registered thresholds, no deployment without evidence.

Live Demos

AI Lab

• Python + Node.js + Claude Code in individual cloud accounts
• Each student gets their own environment — nothing to install
• ai-lab.rice-business.org
• Log in: test_student / jgsbai

XYZ Corp

• Simulated enterprise data with Claude + Python
• Students query, analyze, and visualize in a realistic setting
• xyzcorp.rice-business.org
• Log in: test_student / jgsbai

Presentation Examiner

• AI assessment and feedback for student presentations
• AI-generated questions and follow-ups based on slide content
• presentation.rice-business.org
• Student demo link

Three live tools we’ve built. AI Lab gives each student a cloud environment with Python, Node.js, and Claude Code. XYZ Corp is a simulated enterprise dataset for data analysis exercises. The Presentation Examiner does AI-driven oral assessment. All three are live — let me show you.

Who Wants to Pilot This?

I’m looking for 2-3 faculty across different disciplines to pilot this in Spring 2027. You keep full control of your course. We add the AI examiner as one assessment component and compare results. Who’s interested?