Week 04
AI-Assisted Analysis

API209: Summer Math Camp

Rony Rodrigo Maximiliano Rodriguez-Ramirez
rrodriguezramirez@g.harvard.edu

Harvard University

September 2, 2026

AI for Fall Semester Work

The point of this week

Use AI to strengthen judgment, speed, and verification

  • faster debugging
  • faster exploration
  • better documentation
  • stronger audit habits

What good AI use looks like

  1. Define the task before prompting.
  2. Ask for one bounded output.
  3. Run and inspect the result yourself.
  4. Verify one important claim manually.
  5. Document how AI changed your workflow.

Where AI helps most in API209

Good uses

  • debugging code
  • translating plain English into dplyr steps
  • drafting a plot specification
  • generating a Quarto skeleton
  • extracting structure from messy text

Weak uses

  • writing a whole assignment for you
  • inventing citations or sources
  • interpreting a model you have not checked
  • making policy claims from one quick graph
  • hiding sloppy workflow behind fluent prose

Two tools to know

Codex

  • OpenAI’s coding agent for terminal, IDE, and app workflows
  • useful for repo navigation, code edits, test runs, and debugging
  • strongest when you give it a clear task and verify the output

Claude Code

  • Anthropic’s terminal coding tool
  • useful for feature work, bug fixing, code explanation, and automation
  • strongest when you provide context and keep the task scoped

A practical prompt pattern

I am working in R on a small policy-data assignment.
I have a data frame called `pisa_long`.
Write a short `dplyr` pipeline that computes the weighted mean
score by country and test, using `stu_wgt`.
Return only the code and one sentence explaining the logic.

Prompting for better outputs

  • name the language
  • name the objects
  • name the expected output
  • name the constraints
  • ask for a table, checklist, or code block when possible

Verification is the real skill

The best students are not the ones who trust AI fastest.

They are the ones who can verify AI output quickly.

Mini exercise 1

You have an error in a ggplot() call.

Ask Codex or Claude Code to:

  • explain the error,
  • propose the smallest fix,
  • tell you what object to inspect next.

Then test whether the fix actually works.

Mini exercise 2

Take a rough chart and ask the tool to critique:

  • the title,
  • the color choices,
  • the axis labels,
  • the main interpretive risk.

Keep only the advice that improves clarity.

Mini exercise 3

Ask for a Quarto template that includes:

  • setup
  • data section
  • analysis section
  • figure section
  • verification section

Then adapt it to your own assignment rather than using it unchanged.

Ground rules

  • AI is allowed for help, not for blind delegation.
  • Every meaningful result still needs human review.
  • If AI materially shaped your work, disclose it.
  • Reproducibility matters more now, not less.

Carry this into the fall

plan → ask → run → verify → document