Sep 25 · Week 2 materials
Everything from the first session — slides, handout, what was covered, and how the two hours ran.
Course materials are also mirrored in the shared Google Drive folder.
These panels are closed. Click a heading to open it.
Slides and handout deck · printable handout · web version
The deck and a two-page reference sheet. The handout is the one to print or keep open during the demos — it carries the framework the four levels are built on.
What the lecture covers eight parts
The first half explains how AI models generate outputs and why plausible answers can still be wrong. The second half shows how to build more reliable agent workflows with structured tool interfaces, MCP, permissions, and scoped subagents — ending with four hands-on demos.
| Part | What it covers |
|---|---|
| Models and scientific workflows | What models do, how they differ from algorithms and workflows, and where they appear in research. |
| How LLMs generate | Tokens, next-token prediction, context, learned priors, and why plausible output is not always correct. |
| Attention and Transformers | How attention evolved into self-attention and enabled modern long-context language models. |
| Why coding agents fail | Context dilution, competing examples, ambiguous parameters, and silent software errors. |
| Structured tool interfaces | Capability descriptions, schemas, validation, and structured inputs and outputs. |
| Model Context Protocol | How MCP exposes selected research capabilities through a common interface. |
| Permissions and scoped subagents | Restricting direct access, narrowing tool sets, and delegating work to scoped subagents. |
| Live tissue-analysis demos | Progressing from a raw repository to MCP, permissions, and scoped subagent delegation. |
Central message. Using an agent means asking an AI system to help do the work — "here is my project, please figure out how it works and analyse my data."
Designing an agent means deciding what the AI should and should not have to figure out. Instead of asking it to rediscover everything from source code, we define the workflow explicitly: inputs, allowed parameters, outputs, validation rules, and permissions.
How the session runs Sep 25 timetable
Two hours: about 70 minutes of lecture and live demos, then 50 minutes of setup, group work and close.
| Time | What happens |
|---|---|
| 2:30 | Open |
| 2:35 | AI models — how they generate, and why a plausible answer can still be wrong |
| 3:15 | Break, and a setup check: run claude, then /status should show the NRT organization |
| 3:25 | Structured tool interfaces, MCP, permissions, and scoped subagents |
| 3:35 | Four demo moments, live — one from each level |
| 3:50 | Group work — setup if you still need it, otherwise Demo 1 then Demo 2 |
| 4:20 | Close |
| 4:30 | End |
You are not expected to finish all four demos in the room. Demos 3 and 4 are yours to work through afterwards — every guide on this page is written to be followed on your own.
Oct 2 · Week 3 materials
A hands-on session: Claude and Claude Code side by side, Markdown and the handful of terminal commands you need, the course project, the practice demos, and a project brainstorm. The goal is to leave with a working setup and one project idea of your own.
Please complete the setup before Friday by working through the “Before you arrive” section on this page. The session will move together as a group, so having the software installed and ready in advance will make it much easier to follow along.
Do not buy a plan. Create a Claude account only — the AIDMM NRT provides a Claude Team seat. Send the email address you want registered to Yanrui Ma. If you already have a working Pro-level Claude Code account, skip the account and license steps.
These panels are closed. Click a heading to open it.
Follow-along sheet print it · bring it Friday
Four pages, numbered in the order the session runs. Every command and prompt you will type is on it, with blank lines to write down what you got. Print it or keep it open beside the terminal — this is the one thing to have in front of you on Friday.
Before class — the checklist do this first
Five things, in order. The first three are the ones that matter.
- Create a Claude account — no plan. Then email mayanrui@msu.edu the address to register for the NRT Team seat.
- Set up a terminal. Field Guide Part 4 (slides 28–34), or use VS Code — see Choose a setup path below.
- Install Claude Code and log in. Field Guide Part 8
(slides 62–65), through First run and login. Then run
claudeand type/status— it should show the NRT organization once your seat is active. Stop there. - Read Parts 1–3 of the Field Guide. What an agent is, how access and limits work, and Markdown — the format every agent instruction file is written in.
- Optional. Try Your first ten minutes (slide 68) against the
practice project; unzip
survey-study.zipfirst.
If something fails, come anyway — bring the error message. Nothing in Friday's session is lost by watching the first half from a neighbour's screen.
How the session runs Oct 2 timetable
Two hours, almost all of it hands-on. Each block is demonstrated first, then you follow along.
| Time | Activity | What you should get out of it |
|---|---|---|
| 2:30–2:45 | Claude basics Instructor-led + follow along |
Claude vs. Claude Code — what each environment is for. Open or install as needed, sign in, and learn how to talk to Claude and respond to approval prompts. |
| 2:45–3:05 | Markdown + terminal basics Demo + group |
Read a Markdown file. Open a terminal and practise only what class needs: pwd, ls, cd, cd .., and moving around paths. |
| 3:05–3:20 | Open the course project Demo + group |
Navigate to the course folder, launch Claude Code, confirm it can see the project files, and get one simple prompt to work end to end. |
| 3:20–4:00 | Practice demos Demo + group |
Repository-as-context against MCP: how the agent infers, how permission-aware tools behave, and how to define a scoped agent. |
| 4:00–4:30 | Brainstorm project ideas Group |
Find one workflow you already run that could become an agent-callable capability, and sketch it. |
Advice — how to get the most out of Friday read this
Prepare candidate workflow in mind. The last half hour is the part that outlasts the workshop. Think of a step you already run regularly: the inputs are files, the output is a number, a table or a figure, and you can say what a wrong answer would look like. That last point is the one worth bringing — being able to state the check is what turns a script into a capability worth exposing.
Keep the handout open in a second tab. The vocabulary — model, workflow, tool, schema, permission, scoped subagent — is doing real work in every demo.
Expect not to finish, and do not let that stall you. Every guide on this site is written to be followed alone afterwards. If you fall behind during a block, stop typing and watch; the files are all still there on Saturday.
Pair up if your setup breaks. Two people on one working laptop learn more than one person debugging an install for forty minutes.
Oct 9 · From chatbot to coworker
An introduction to the Claude landscape and the four ways you can work with it — chat, Cowork, Claude Code in the desktop app, and Claude Code in the terminal — told through one physics problem solved four times over. What each one can reach, what it remembers between sessions, and how to pick the right one for the job in front of you.
Led by Prof. Michael Murillo, Michigan State University.
The idea to carry forward: the model stays fixed — what changes is the scaffolding around it. Reach decides what an agent can do; memory — files, rules, skills — decides whether it does it the way you need. That distinction is what the Oct 16 session builds on.
Oct 16 · Pattern design intro
Designing an agent for a task you already do. Four patterns to apply to any workflow — who does each step, how much to build, where it runs, and how you would know it went wrong — then a live demo that turns one real task into a skill you can run again.
Design the workflow first. The agent is only one of the boxes. Most steps in a real task belong to you, to a script, or to a check. Friday is about deciding which box is which — and “this does not need an agent” is a perfectly good conclusion.
Bring one small, ordinary task you already do and the five short answers about it: what you do, how you do it now step by step, a sketch of that, what goes in and out of each step — who does it, and whether it belongs in chat, Cowork or Claude Code — and what could go wrong without you noticing. Half an hour is plenty, and a photo of a hand-drawn workflow is a perfect answer.
Submit your solution here: forms.gle/EDDonqW83ZDZKEtz5
Share what you built — your own task or the ADS demo. A link to
your folder, your SKILL.md, or a short note on what worked and what did not. An
answer that concludes “this did not need an agent”, with the reasoning, is
complete.
In a group project? The task you describe here is a good
candidate for the assignment of one problem solved four or five ways — and questions 1, 2
and 5 are most of what a seed.md needs. Bring both.
Also bring one real file that your task starts from — a real spreadsheet, document or export, messy as it is, with names changed if it is sensitive. Made-up files are tidy and real ones are not, and everything interesting lives in that difference.
Before you arrive
Everything referenced in the preparation email is available below. Because Friday is hands-on, please complete the setup beforehand.
1 · Get your Claude seat
Skip this if you already have a Pro-level Claude Code account running.
- Register at claude.com using your MSU or EGR email address — a university address is required to be added to the Team plan.
- Email that address to Yanrui Ma (mayanrui@msu.edu) to request your seat. You will receive an email invitation.
Do not buy a plan. The AIDMM NRT covers the cost of your Claude Team seat for the duration of the workshop. If you have not requested a seat yet, email Yanrui now rather than waiting — the invitation has to be issued before Friday.
Your seat is working when /status inside Claude Code shows the
NRT organization. That is the one check worth doing before you arrive.
One habit worth starting with: Claude Code asks before it acts, and that is the
default. shift+tab cycles to modes that stop asking — leave it alone on Friday.
Demo 3 is entirely about watching permissions refuse things, and in a bypass mode nothing
refuses.
2 · Choose a setup path
Plain terminal
macOS: Part 4 of the Field Guide. Terminal is already installed, under Applications → Utilities.
Windows: use PowerShell (Start → PowerShell).
Fine if you are comfortable at a command line and not using VS Code.
VS Code
Follow Parts 1–4 of the VS Code setup guide: extensions, the Python interpreter, and running a script.
Windows users should take this path. Friday's demo uses VS Code.
Inside VS Code, ⌘ J on macOS or Ctrl J on
Windows opens the terminal.
3 · Install Claude Code
Work through Part 8 of the Field Guide, from Step 0 · Before you install up to and including Step 2 · First run and login (slides 63–65). If you prefer the official instructions, searching for "install Claude Code" gets you there too.
Stop once you have logged in and seen the prompt. Everything after that we do together.
4 · Read the background — Parts 1–3
Three short parts of the Field Guide, worth reading even if your install already works:
- Part 1 — What it is. What an agent actually is, and how it differs from a chatbot.
- Part 2 — Access, cost and limits. How seats, usage and policy work.
- Part 3 — Markdown. The format every agent instruction file is written in — and the one you will be editing on Friday.
5 · Optional, if you have time
Try Your first ten minutes (Field Guide, slide 68) against the practice project.
Unzip survey-study.zip somewhere sensible first:
mkdir -p ~/projects && cd ~/projects
unzip ~/Downloads/survey-study.zip
cd survey-study
git init && git add -A && git commit -m "Starting point"The study's notes and its data deliberately disagree with each other. Finding the differences is the exercise — and it is a much better first conversation with an agent than asking it to write something from scratch.
6 · The demo package
Download Demo 1 above and unzip it. Inside the folder, create the environment the guide uses:
python3 -m venv .demo_venv
source .demo_venv/bin/activate
pip install -r requirements.txtOn Windows the activate line is .demo_venv\Scripts\activate. The
Demo 1 guide walks through this and everything after it.
If something fails, bring it to the session rather than fixing it alone. If one laptop is stuck, three more usually are, and it is quicker to sort out together.
Nothing here needs a GPU or an API key
The demos use small synthetic slides in a plain-text image format, so they run on any laptop. The analysis is a stand-in for a real histology pipeline — the failure modes are real, the pixels are not.