MGMT 638, Fall 2026
Tue Aug 25, 2026
The tool
Generative AI as it is actually used at work: to write and run code.
The subject
Quantitative investing. Predicting stock returns, and backtesting strategies.
Gen AI signals
Generative AI + news → trading signals
| Component | Weight |
|---|---|
| Group presentations | 60% |
| Class participation | 40% |
Group size = 2 or 3. Each group presents twice. Three presentations each class session beginning next week
Class participation includes cold calls and share screen.
What runs
Claude Code, running GLM-4.7.
What is installed
Python, JavaScript, the analysis libraries, and a set of Claude Code plugins.
What you see
A file browser, a terminal, and a browser view for anything you build.
All of it runs in the cloud and is billed to the school.
lab638.kerryback.com — any browser, nothing to install@rice.eduS0First load takes a moment while your machine starts. If you are asked to sign in again later in the term, nothing is wrong — sessions do not survive a restart of the server.
Right panel
Claude Code. The prompt window is at the bottom.
Left panel
File browser by default. Terminal and View are there too — ignore them today.
Three dots, upper left
Upload from your laptop, download to it.
| Refresh | The file browser does not update itself. When GLM makes a file and you cannot see it, hit Refresh. |
| No Office | Double-clicking a .xlsx or .pptx does nothing. Office is not installed. |
| Download to view | Bring Office documents to your laptop and open them there. |
Most of the prompt engineering advice from a few years ago is now either useless or actively harmful.
Say what you want, not how to do it
“I want to compare revenue across countries — how should we show it?”
beats
“Make a bar chart, sorted descending, gridlines off, 12pt labels.”
If it does not work, try again
Rephrasing is cheap. Treat it as an infinitely patient colleague with a poor memory.
Say: plan how to do this before you start.
Why it works — knowledge
Planning pulls the relevant considerations into the conversation. It then answers in light of what it just retrieved.
Why it works — subtasks
It breaks the work into steps you can check one at a time. Failures stay small and local instead of surfacing at the very end.
Ambiguity does not produce a question. It produces a confident answer to a question you did not ask.
Tell GLM, before it starts
Tell me what you are about to do and what you are assuming.
Correct it in one sentence, then say go ahead. This is the cheapest habit in the course and it will save you the most time.
Suppose it told you last week that method A beats method B. Today you ask it to do the task.
What it will do
If “A beats B” is not in this conversation, and B appeared more often in its training data, it will probably use B.
Why
Knowledge in the weights is not knowledge in use. Something has to pull it into the conversation before it can act on it.
Prediction also means momentum. Once it has started down a path, continuing that path is the likeliest next step.
You will watch it hit a roadblock, find a way around, hit another, find another way around — tenaciously, exhaustively. What it rarely does is go back and ask whether a different path was better all along.
Restate the goal, not the obstacle. That gives it permission to abandon the path instead of defending it.
| Skip | Why |
|---|---|
| “Think step by step” | Reasoning models already do, in hidden tokens |
| “Take a deep breath” | Measurably useless now |
| “CRITICAL! NEVER EVER …” | Overtriggers, and gets worse results |
| Stacked personas | “You are an expert X” can help tone; three of them does not |
Direct task framing wins. Say what you want, say what done looks like, say what to avoid.
You
↔︎
Chatbot
↔︎
The model
The model reads text and writes text. That is the entire capability.
It cannot open a file, run a calculation, look up today’s price, or check whether what it just told you is true.
Code execution
↕
You
↔︎
Agent
↔︎
The model
↕
Other tools
The model never runs any of it — it only emits requests.
The model emits requests; the agent executes them. Those step-3 instructions are the tool harness, and some are far better than others.
Almost everything useful the model does, it does by writing code and running it, including terminal commands.
On our AI lab, GLM will write code, save it to a file (a tool use), then run it with a terminal command (another tool use), read the output (another tool use), and revise and repeat as necessary.
All of this is orchestrated by the Claude Code “harness.”
Excel workbooks with live formulas
Charts and figures
Word and PowerPoint files
Data analysis and statistics
Web apps and dashboards
Nov 2022
ChatGPT opens to the public
Mar 2023
GPT-4 and Claude — the chatbot year
Nov 2024
Model Context Protocol — agents get tools
Feb 2025
Claude Code — the agent reaches your files
Oct 2025
Agent Skills — expertise you can package
2026
Cowork and desktop agents
OpenAI co-founder, former Tesla AI chief, now at Anthropic
October 2025
“They just don’t work. The industry 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.”
February 2026
“Coding agents basically didn’t work before December. I don’t think I’ve typed like a line of code probably since December.”
A commit is a snapshot of the whole folder, dated and labelled. Nothing is ever overwritten — it is added to.
Every commit is a point you can go back to, and Git can tell you what changed between any two of them, in which file, and on which line.
Git is not an AI tool. It is twenty years older than any of this. GLM simply knows how to drive it.
Git
On your machine. The history lives in the folder itself — no network, no account, yours alone.
GitHub
A hosted copy of that same history for backup and sharing.
github.com, and note your usernamegh auth login and answer: GitHub.com, HTTPS, yes to authenticating Git, browserTell GLM — one person in the group, once
Make a folder in my workspace called [whatever you want to call it] and make it a git repository.
Add a .gitignore for csv, parquet, and .env files.
Commit, then create a private GitHub repo of the same name and push to it.
On github.com, in the new repository
Tell GLM
Clone the GitHub repo [username]/[repo name] into my workspace.
username should be the username of the group member who created it. repo name is the name they used for the folder
Tell GLM, at the start of every session
Pull the latest from GitHub before we start.
Then, at the end of the session
Commit what we did and push it.
| What you see | What to say |
|---|---|
rejected --- non-fast-forward |
Someone pushed while you were working. “Pull their work, merge it, then push.” |
CONFLICT: automatic merge failed |
You both edited the same lines. “Show me both versions and wait for me to choose.” |
Refuse these if asked: git push --force and git reset --hard
MGMT 638 · Gen AI and Quantitative Investments