NUS Business School

AI Tools for Business Research · 4 & 5 August 2026

Materials

Everything used across the two days: the setup you do first, the five decks in running order, the three practice kits, and the resources worth keeping.

Start here

Budget 30 to 45 minutes for the environment. Participants did this before day one so that no teaching hour was lost to installing software; if you are working through the decks on your own, do it first for the same reason. The full walkthrough, including Windows and troubleshooting, is in the setup guide.

  • Two accounts — a Claude Code subscription (or API access), and GitHub.
  • A terminal — already installed on macOS; on Windows, install WSL with wsl --install.
  • Python 3.10 or later — check with python3 --version.
  • Claude Code — npm install -g @anthropic-ai/claude-code, then run claude once and log in.
  • Obsidian — free, no account needed, for Session 3.

Then check it works

Make a scratch folder, start the agent, and ask it for something small:

> Create a file called hello.py that prints "ready for the workshop",
  then run it.
Claude asks permission, writes the file, runs it, prints the message. If that happened, you are set up correctly — quit with /exit.

If it does not work, email jiaqishao@u.nus.edu with the exact error text.

Slides

Five decks, in running order — each the deck used live in the room. Sessions 1B, 2, 3, and 4 are built with the house Beamer style, so the covers below are also a fair preview of what is inside.

Session Deck Pages Size
1A From ledgers to agents: the long automation of economic research Day 1 morning 74 20 MB
1B Getting started with Claude Code: coding basics and agentic systems Day 1 late morning 60 242 KB
2 Vibe coding with Claude Code: two tasks, four rounds, one idea Day 1 afternoon 112 995 KB
3 Wiki LLM: turn scattered papers into knowledge you can find and check Day 2 morning 44 3.2 MB
4 From filings to a structured database: sixteen steps, you type all of them Day 2 afternoon 51 232 KB

Practices

Three hands-on blocks, twenty minutes each, all on the same paper. They compound: A gets you a verified reading, B turns it into two deliverables, C turns it into knowledge that keeps paying back. Session 4 is a fourth hands-on block, but it is not a twenty-minute practice — the whole three hours is the build. Full instructions are in the practice guide.

The paper was handed out in the room. All three practices run on one field-experiment manuscript that is unpublished, so it is not distributed here. Everything else you need — the kit, the instructions, the slides — is on this page.

Install the kit first

Everything runs out of the setup kit — a project memory file, a faithfulness rule, a house Beamer style, the paper-to-slides skill, and a referee subagent. It is adapted from Pedro Sant'Anna's public academic workflow. From the folder holding the paper:

$ mkdir -p .claude
$ cp claude-setup/CLAUDE.md ./CLAUDE.md
$ cp -r claude-setup/{rules,skills,agents} .claude/
$ claude
Then check it loaded: /memory shows CLAUDE.md, and /agents lists referee.

A · Set up, then understand the paper

Session 1B · 20 min · PA_practice.pdf

You have just met the agent; now put it on a real paper. You produce a working setup and a verified one-paragraph overview. Set up, five minutes; orient, eight; verify, six. The habit you keep for the rest of the workshop: never accept a number without naming the table it came from.

B · Paper to talk, then to referee

Session 2 · 20 min · PB_practice.pdf

Two finished artifacts from one paper: present it fairly, then referee it fairly. You produce a faithful talk deck and a referee report, using plan mode, the paper-to-slides skill, and the referee subagent. The instructor's demo is the target — yours need not match slide for slide, only fact for fact.

> Plan a faithful conference talk from @RCT_paper.md using the
  paper-to-slides skill. Show me the slide outline before drafting.
Plan mode is a read-only gate — Shift+Tab, and you see the outline before a single file is written.

C · Make the paper compound

Session 3 · 20 min · PC_practice.pdf

A talk is used once; a wiki keeps paying back every time you return to the paper. You produce a mini LLM wiki — an index, a few pages, a log, then a lint pass. The raw source stays read-only; the agent maintains the knowledge pages. Next time you query the wiki instead of re-reading forty pages.

Going further

Tools and documentation

Data sources used or mentioned in the sessions

Papers worth reading alongside

  • Gentzkow, Kelly & Taddy (2019). Text as Data. Journal of Economic Literature 57, 535–574.
  • Korinek (2023). Generative AI for Economic Research: Use Cases and Implications for Economists. Journal of Economic Literature 61, 1281–1317.
  • Roth, Sant'Anna, Bilinski & Poe (2023). What's Trending in Difference-in-Differences? Journal of Econometrics 235, 2218–2244.
  • Hassan, Hollander, van Lent & Tahoun (2019). Firm-Level Political Risk: Measurement and Effects. Quarterly Journal of Economics 134, 2135–2202.

Journal policies on disclosing AI use