AI Bootcamp: Tools for Economic Research

Claude Code, Codex & the Research Pipeline

Julian Hinz

Kiel Institute for the World Economy

Paula Jacobs

Kiel Institute for the World Economy

2026-04-16

Watch this.

Open Terminal on the projector. Briefly normalize: “This is Terminal — a place where you type commands instead of clicking. That’s all.”

Navigate to the demo project folder. Type:

“Load the BACI bilateral trade data and CEPII distances. Plot Germany’s 2022 exports to all destinations against bilateral distance, log-log scale, with a fitted line. Label the slope and key partners (US, China, France, Poland). Publication-ready ggplot.”

Wait for the gravity plot to appear. Then:

“That’s the gravity law for Germany — slope of about -1. I didn’t write a single line of code.”

No slides during this — the figure on the terminal is the slide.

Why This Matters

Research Is a Pipeline

Walk through the 8 stages briefly: results reshape the question, new data changes the analysis, reviewer feedback sends you back to an earlier stage. Then: “Which of these stages eats the most of your time?” Take a few answers from the audience before moving on.

Source: Paul Goldsmith-Pinkham, Research in the Time of AI

The Fear: AI Collapses the Pipeline

This is the caricature. It’s worth stating explicitly because it’s what people are afraid of. “No design iteration. No critical reading.” Next slide: the more accurate picture.

Source: Paul Goldsmith-Pinkham, Research in the Time of AI

The Reality: AI Compresses the Pipeline

This is THE central insight. “You move from an idea to a first result in an afternoon instead of two weeks. You iterate five times where you used to iterate once.” The win is not “AI writes your paper.” The win is “AI lets you iterate more, and faster, so you find better questions and better answers.”

Source: Paul Goldsmith-Pinkham, Research in the Time of AI

The “Slop” Anxiety

• More papers written with less care
• Less friction → more output (COVID spike)
• AI could make that permanent
• But replication got cheaper too
• Cunningham (CC36): honest public self-correction

Goldsmith-Pinkham · Cunningham CC series

The Job Anxiety

What LLMs replicate

• Writing code
• Cleaning data
• Implementing estimators
• Producing first drafts
• Formatting tables

→ Execution skills — depreciate fast

What they don’t

• Which question matters
• Credible identification design
• Spotting when data lies
• Iterating until it works

→ Research judgment — appreciates

Even If You Never Run a Regression

• Summarize 30 papers in minutes
• Regression table → policy brief
• Extract themes from interviews
• Draft R&R responses
• Triage inbox by VIP/topic

claudeblattman.com

What These Tools Are

A Note Before We Start

• You don’t need to become a terminal user (there are app versions of this too)
• Some will use these tools directly
• Others will supervise an RA who does
• Today: understand what’s possible

How LLMs Work

• Trained on vast amounts of text
• Idea: Predict the most likely next word
• No memory between separate calls
• No internet access by default
• Pattern matching, not understanding
• Prompt: Your input that gets encoded

Keep this very simple. The audience includes people who have never thought about how AI works. The key insight: it’s a text prediction machine, not a database, not a search engine, not a thinking being. It has no persistent memory — every session starts from zero.

The “Chat”

Everything the model can see when it generates the next response. System prompt, CLAUDE.md, conversation so far, tool calls + outputs, files read — and then whatever free space is left. 1M tokens hard cutoff. The model re-reads ALL of this every time it responds.

From Chat to Agent

• Same LLM, new scaffolding (“harness”)
• Can read your files
• Can write and run code
• Can iterate on errors
• R, Python, Stata, LaTeX — anything

This is the conceptual leap. The model is identical — what changes is what it’s allowed to do. Claude.ai = Level 0–1. Claude Code = Level 3. Same brain, different hands.

What Is a Terminal?

A text interface to your computer. You already know one: the Stata command window, the R console, a Jupyter cell.

$ pwd                        # Where am I?
/Users/julian/Repositories/KITE-PB-China-Africa-Tariffs

$ ls                         # What's here?
code/   input/   output/   CLAUDE.md   Makefile   README.md

$ claude                     # Start the agent

Terminal basics guide

CLAUDE.md — The Briefing Memo

Write it once. Claude reads it every session.

# China–Africa Zero-Tariff Policy Brief

## Goal
Quantify the impact of China eliminating all tariffs on imports
from 53 African countries, using the KITE trade model (GTAP 11).

## Data
- input/baci/ — BACI bilateral trade flows, HS6, cols: t, i, j, k, v, q
- input/initial_conditions/ — GTAP 11 initial conditions (.rds)
- input/metadata/countrygroups.R — AFRICA53, AFRICA_LDC, EU27

## Stack
- R with data.table, magrittr, ggplot2
- Figures: theme_minimal(), Kiel cream #F5F1E7, blue #194ABB
- snake_case for variables, SCREAMING_SNAKE for constants

## Rules
- Always copy initial conditions before modifying (copy-first pattern)
- Save every figure as PNG + PDF pair
- Never edit docs_policy_brief/ — edit output/policy_brief.tex instead
- Commit after every completed step

guides/claude-md.qmd

When the Window Fills

guides/context-windows.qmd

What About My Data?

The Dropbox rule

If you wouldn’t put it on Dropbox, don’t let Claude read it.

• Files stay on your machine
• API content not used for training
• But what Claude reads is sent to the API
• .claudeignore to exclude sensitive folders
• Use aggregated extracts for confidential data

guides/privacy.qmd

Live Demo

Where This Sits

Level	What it is	Example
0–1	Browser chat, copy-paste	ChatGPT, Claude.ai
2	IDE-integrated agent	Cursor, GitHub Copilot
3	Terminal agent, full file access	Claude Code · Codex CLI
4	Agent + external connectors	Gmail, Zotero, databases
5	Autonomous sub-agents	Long-running, self-directed

You just saw Level 3. ~$20/mo Pro · ~$100/mo Max

Adapted from Goldsmith-Pinkham · guides/costs.qmd

Verify Everything

• LLMs produce plausible but sometimes wrong output
• Check observation counts after merges
• Check coefficient signs and magnitudes
• Check assumptions (SEs, sample restrictions)
• You are the expert. Claude is the RA.

This is the single most important trust-building slide. Economists in the room need to hear that you take verification seriously. Mention: “In the demo, I checked each result before moving to the next step. That’s not optional.”

BREAK

15 minutes

Nuvolos login on screen · companion website QR code

Survey Results

Two groups in the room

“We asked you before the bootcamp. Here’s what you told us.”

Half the room is Level 0 — and that’s overwhelmingly the senior researchers. PhDs are spread from 0 to 5. Mean ladder: 1.9 for PhDs, 0.5 for seniors. 73% of seniors have never gone beyond browser chat.

88% of PhDs use AI regularly or more. 47% of seniors only occasionally. Most people are on free tiers or one subscription — 60% of seniors pay nothing. ChatGPT dominates (24/32). Claude Code already known to 10, but almost all PhDs.

This isn’t surprising — but it shapes what we’ll do this afternoon.

What you want vs what you know

The orange-heavy bars mean low scores, the blue-heavy bars mean high scores. Confidence is on the left, want-to-learn on the right. The number in between is the gap.

Look at multi-step tasks: highest demand, lowest confidence. Gap of +2.5. That’s the gap we’re here to close.

Reproducibility is second (+2.1). You want to trust AI output but you don’t know how yet. We’ll cover that.

Prompting, writing, lit search — you already feel OK about these. Good. We won’t spend much time there.

Different starting points

When we split by career stage, the picture sharpens.

PhDs (blue circles) already feel confident on the basics — prompting, coding, lit search. They want to go deeper: multi-step tasks, cost-conscious use, reproducibility.

Seniors (orange squares) have gaps everywhere, but their biggest want is the same: multi-step tasks. They know that’s where the leverage is.

“That’s why we have two tracks. Not because one is better — because you’re starting from different places and want different things.”

“Let’s close the gap. Here’s how the next hour works.”

Hands-On

First — Everyone Types the Same Thing

Everyone types the same first prompt. Nobody moves on until it works.

“What files are in this folder? Summarize the data.”

Buddy system: comfortable with the terminal? Pair with someone who isn’t. You type, they steer.

Two Tracks

Beginners

• First time with a terminal?
• Short guided prompts
• Each builds on the last
• Confidence first, sophistication later

→ exercises/beginners

Advanced

• Already use Cursor or Copilot?
• Write a CLAUDE.md + plan.md
• Initialize Git, execute, push
• Full workflow in 40 minutes

→ exercises/advanced

R/Git user but new to AI agents? Beginners track is a safe bet. You can switch any time.

Advanced Patterns

CLAUDE.md in Depth

Global for you · Project-level for this project · Both read at session start.

For code projects

• Data sources, column names
• Language, packages, versions
• Naming conventions
• Plotting style and colours
• What NOT to do

For writing projects

• Project context and goal
• Voice, tone, ban list
• Key references, terminology
• Constraints and boundaries
• What NOT to do

guides/claude-md.qmd

The Plan-Review-Revise Loop

80% planning, 20% execution.

1. Plan — steps, files, checks
2. Review — fresh session, skeptical prompt
3. Revise — fix what the reviewer found
4. Execute — one session per task

Same-session review = self-peer-review. New session = honest referee.

Van Horn: 70 plan files, 263 commits in 30 days. Traditional development is 80% coding, 20% planning. With AI agents the ratio flips.

A good plan.md describes the goal, lists files to touch, specifies tests or checks, and leaves a checkbox for every step. Claude reads it, executes step by step, and updates the file as it goes. If something breaks, you start a new session, point at the plan, and pick up where you left off.

Blattman’s key insight: Claude reviewing its own plan in the same conversation is like asking someone to peer-review their own paper. Sycophancy is baked into the same conversation. The fix is a fresh agent with no memory of the previous conversation. New chat. New context. Adversarial prompt.

Familiar territory for economists — this is just referee reports for your own code.

claudeblattman.com

Git Is Not Dropbox

Dropbox

• Syncs files automatically
• Versions by timestamp
• paper_final_v2_REAL_final.docx
• No history of why
• One editor at a time

Git

• You choose when to take a snapshot
• Every snapshot explains why
• Clear, queryable decision history
• Multiple people, merged cleanly
• Go back to any previous state

guides/git-basics.qmd

You Don’t Need Git Commands

Git has ~150 commands. Just ask in plain English:

“Commit what we just did with a descriptive message.”

“Push this to GitHub.”

“What has changed since yesterday?”

You stay in the what. Claude handles the how.

guides/git-basics.qmd

Multi-Session Workflows

State lives in files, not in the chat.

Session 1    Clean data            → output/cleaned.csv
Session 2    Read cleaned.csv      → run regressions → tables
Session 3    Read tables           → make figures
Session 4    Read figures + tables → draft paper section

CLAUDE.md persists. Results persist. The chat does not need to.

Five Ways to Manage Context

Fresh session = your brief only. Compact = lossy summary. Subagent = delegate a task, only the result returns. Rewind = keep the good prefix, cut the bad tail. Continue = everything stays. Move left for cheaper context, right for more continuity.

Correcting vs Rewinding

Correcting: each failed attempt stays in context, polluting the window. Rewinding (Esc Esc): you keep only the prefix and the informed correction. Result: cleaner context, better output.

Subagents

The parent spawns a subagent for an exploration task. The subagent does 20 file reads, 12 greps, 3 dead ends — but only the final report returns to the parent. The exploration noise is garbage-collected. Parent context stays light.

Where This Is Going

• Now: Codex CLI (OpenAI’s terminal agent) — good as a second opinion
• Now: MCPs — connectors to Gmail, Zotero, databases
• Emerging: Plan-and-execute splits across models
• Emerging: Autonomous agents running for hours unsupervised

guides/costs.qmd

Wrap-Up

Three Things to Remember

1. AI compresses the pipeline. It doesn’t replace the researcher.

2. CLAUDE.md is your most important file. Brief the AI like a new RA.

3. Start small. One prompt, one task, one session.

Your Next Step

• This week: Read Goldsmith-Pinkham’s Getting Started with Claude Code and following posts
• This month: Write a CLAUDE.md for your main project

… and follow-up AI Lunches / Workshops / Bootcamps?

The Companion Website Has Everything

• These slides
• Setup for Nuvolos and local install (Mac + Windows)
• Both hands-on tracks with full prompts and tips
• Deep guides: terminal, CLAUDE.md, Git, privacy, costs, context windows
• Links to further resources