Working with AI in this course

On this page

1. Start from what you want to build
2. Let it see the data itself
3. Read it before you accept it
4. How you know it is right
5. What to do when it is wrong
6. Before you hand it in
7. What gets rewarded

You are expected to use AI here. It is the subject of the course. The rules are in the syllabus, under AI Policy, including the four levels that every assignment is labeled with. This page is about the doing.

Related pages:

• Your AI Tools and Babson Key: the Babson key, setting up the agent in marimo and in VS Code, your agent's rules, and other AI tools
• Python Handbook, Chapter D: what a request to a model contains, tools, and the agent loop
• Recommended Reading, under Working with AI: further reading, none of it required

Start from what you want to build#

Before you ask for code, tell the agent what you are making and who will use it, such as "a tool my friend can use to compare two mortgage offers". Then ask what it would build it with: which data source, which library, which API. It knows of far more than this course covers. You choose among what it offers, and you can state why.

When it proposes something you have not used, ask for the smallest version that runs, then build from there.

Let it see the data itself#

An agent that cannot see your dataframe has to guess what your columns are called, what type they are, and how many rows came back. It guesses fluently. The code it writes looks right and runs, and then a column that is text where you thought it was a date sorts 01/01/2016 before 07/13/2012.

So the agent this course uses reads your running notebook, in Agent mode, whether you work in marimo's panel in your browser or in VS Code's Chat. It can run a cell and read the output. When you are stuck, the fix is usually to let it look:

• run df.dtypes and df.shape, so the answer sits in a cell's output. marimo's panel reads your running notebook itself. In VS Code's Chat, type # and pick the variable, or use Add Cell Output to Chat on the cell
• name the table and the columns
• paste the actual error, all of it, including the line numbers

In VS Code, the agent can also read the errors in the Problems panel and the output of a command it ran itself. For an error anywhere else, paste the full text or attach a screenshot.

Anthropic's own prompting reference files this under its own heading, Add context to improve performance. The rest of that page is written for developers building on the API, so read the one section and stop.

Read it before you accept it#

In marimo's panel, a change arrives with Keep cell and Discard cell, and nothing reaches your notebook until you press one. In VS Code's Chat the agent adds and runs the cell straight away, so read it before you save the file. Your AI Tools and Babson Key has the details for both. Keeping a change you have not read is how a wrong number gets into a memo with your name on it.

Reading does not mean understanding every character. It means being able to explain what the code does, and to name the one line you would check first if the answer looked strange.

How you know it is right#

Every memo must carry a section headed How I know these numbers are right, with at least one check a reader could run: the same figure computed a second way, a row count before and after a join, a total reconciled against a source. It has its own line on the rubric. A paragraph saying you feel confident does not count.

What to do when it is wrong#

1. Read the error. All of it. The last line names the problem and the line above it names the place.
2. Give the error back to the agent, unedited. Do not summarize it.
3. If the second attempt fails the same way, stop and state what you actually want in one sentence. Two failed attempts usually means it is answering a different question.
4. If the third fails, ask in the room. An hour stuck alone is a question worth asking out loud, and someone else has the same one.

Before you hand it in#

• You can walk through every line if asked on the spot
• It answers the question you set
• Somebody outside this course can open it and use it
• In a project, it goes at least one step past the core, in a direction you chose
• It says how you know its numbers are right

What gets rewarded#

The grade rewards work that somebody outside this course can use, and a step past what was asked. In a project, the Going further part is required, in any direction, and an attempt that does not work still earns its rubric line.

Last updated: Sunday, 11/1/2026