Syllabus
Claude Code for Accounting Research: AI Agents for the Empirical Researcher
This is a five-module, hands-on intensive (about three hours per module) plus a closing half-day seminar, built for accounting PhD students and faculty — mixed Stata/Python backgrounds assumed, no prior experience with Claude Code or agentic AI tools assumed. The premise is that agentic AI is research infrastructure, not a chatbot novelty: over five modules you build real accounting pipelines — a free SEC dataset, EDGAR filing text, WRDS data on your own account — each one carried through to a verified, reproducible result rather than a one-off chat answer. Verification discipline runs through every module, not just the last one, because the entire argument for using an agent this capable is that you can trust the parts you’ve actually checked. The closing seminar steps back from the pipelines to ask what any of this means for how accounting research gets produced and for the academic labor market it feeds — you leave the course with working pipelines, a small personal library of skills, and a one-page position on your own comparative advantage that you’re prepared to defend.
Format and audience
Five build modules, roughly three hours each, following the same shape every module: a short lecture, a live demo (including a failure or two, recovered on camera), a lab built on a starter zip with its own verify checklist, and a debrief with an exit ticket. Module 6 breaks that pattern deliberately — a discussion seminar, no lab, built around a framing lecture and three structured questions. The audience is accounting PhD students and faculty; the material assumes comfort with either Stata or Python but no prior exposure to Claude Code, MCP, or any agentic tool.
Module by module
| Module | Title | Core question | Lab deliverable | Checkpoint |
|---|---|---|---|---|
| 1 | Foundations | What kind of tool is this, and how do I stay in control? | Install + a working CLAUDE.md + one verified task |
WRDS/Duo homework assigned |
| 2 | The Empirical Loop (SEC FSDS) | Can I go idea → verified figure, reproducibly? | FSDS pipeline + figure + replication README | Tunnel-status screenshot due |
| 3 | EDGAR Text → Data, + Skills | Can I build a text measure I’d defend to a referee? | 10-K panel + validity protocol + first SKILL.md |
Duo stragglers’ last call |
| 4 | WRDS MCP at Scale | Can Claude work my licensed data, safely, on my own account? | funda pull → Parquet/DuckDB/metadata (+ Stata extract) |
Tunnel teardown verified |
| 5 | Writing, Safety, Verification | What do I delegate, and how do I make my work auditable? | Capstone: a hardened Module 3 or Module 4 pipeline, scored against the rubric below | Module 6 readings distributed |
| 6 | Seminar: AI & the Research Market | What will I still be paid for? | One-page position memo (due one week after) | — |
Each build module follows the same template: 0:00–0:40 lecture, 0:40–1:15 live demo, 1:15–1:25 break, 1:25–2:45 lab, 2:45–3:05 debrief and exit ticket.
Assessment
- Per-module exit tickets (formative, not graded): one concept from the module, one thing that broke, one thing you’d verify before trusting it — plus a module-specific question.
- Lab verify checklists: pass/fail, self-graded, spot-checked by the instructor.
- Module 5 capstone, scored on three equally weighted axes: reproducibility (does it rerun cleanly from the README, is the commit history meaningful), verification (is a paper trail present and cross-checked against the actual commits, are the sanity checks sourced, is every number in the write-up traceable to a script rather than typed by hand), and hygiene (does the credential audit come back clean, are filters documented, is the tunnel torn down).
- Module 6 position memo: one page, “What I will still be paid for in 2036,” due a week after the seminar — graded on defended specificity rather than length: it should cite something concrete from the assigned readings and the course’s own build-module experience, and name one specific investment you’re making in your own comparative advantage.
Course policies
AI use. AI tools are the subject of this course and are permitted throughout it — including on the capstone — with one exception: the argument in the Module 6 position memo has to be your own. Editor-not-rewriter-style assistance (see the Prompts page) is allowed even there; disclose how you used it, in the spirit of the disclosure norms the Module 6 seminar itself discusses.
Data licensing. WRDS-derived data is never shared, posted, or committed to any repository, by the instructor or by students — every pull happens on your own account, under your own institution’s license, and stays there. This isn’t a course rule so much as a restatement of what your WRDS license already requires; the course’s public SEC data (FSDS, EDGAR) has no such restriction and is freely shareable.
Collaboration. Pair debugging during labs is encouraged — talking through a stuck pipeline with a neighbor is the judgment call this course wants you practicing out loud. Graded deliverables (the capstone, the position memo) are individual.
Attribution. The course draws on cited external materials — see the About page for the full attribution note. The pipelines, skills, and written work you produce over the five modules are yours to keep, adapt, and publish.
Prerequisites and cost
You’ll need a laptop with admin rights (macOS or Linux native; Windows via WSL2), a working local Python or R installation before Module 1 (the course’s own examples are in Python, with a Stata bridge added in Module 4), a Claude subscription or API key, a git and GitHub account, and a WRDS account with Duo enrolled — with a hard checkpoint of a working ssh wrds connection by the end of Module 2, so an account or Duo problem can still be fixed before Module 4 needs it. Full install steps, a plain accounting of what a subscription costs at each tier, and the exact WRDS/Duo setup sequence live on the Setup and WRDS MCP Setup pages — read both before Module 1. A pre-course email with the same checklist goes out about a week before the course starts.