Part 38: A Plug for Paul Goldsmith-Pinkham's Markus Academy Series

  • Author/Source: Scott Cunningham (Baylor University), via Substack

  • Original: https://causalinf.substack.com/p/claude-code-38-a-plug-for-paul-goldsmith

  • Key Ideas

  • Cunningham recommends Paul Goldsmith-Pinkham's Markus Academy video series as the best available resource for learning Claude Code for empirical research.
  • Goldsmith-Pinkham is characterized as a rare "centaur" --- a social scientist with deep econometric theory, excellent taste, and genuine computer science fluency --- representing a newer generation of applied economists.
  • The Markus Academy series covers beginner to advanced content: researcher setup, empty-folder-to-figure workflows, EDGAR text-as-data pipelines, and large-scale HMDA data engineering with DuckDB.
  • Cunningham highlights several key themes from the series: "beautiful figures" as high-value output, metadata tables as self-documenting infrastructure, and the collapse of fixed costs for data engineering.

  • The post also discusses the difficulty of teaching Claude Code effectively in standard formats, since the tool is both easy to learn and hard to communicate via traditional lectures.

  • Summary

Cunningham devotes this installment to endorsing Goldsmith-Pinkham's ongoing Markus Academy series on Claude Code for economists. The bulk of the post is a detailed profile of Goldsmith-Pinkham as a scholar --- an applied econometrician (Bartik instruments, contamination bias in OLS), consumer finance researcher, and borderline computer scientist --- situating him among a cohort of millennial economists (Kyle Butts, Pedro Sant'Anna, Andrew Baker, Brantly Callaway) who combine social scientific taste with deep technical skill.

Cunningham walks through all four episodes of the series: (1) getting started with researcher setup, (2) creating figures from empty folders with Claude Code, (3) building structured databases from EDGAR filings using text-as-data approaches, and (4) tackling truly large datasets (70 GB HMDA data, 291 million rows) with DuckDB and parquet. For each episode, he highlights both the pedagogical approach and the substantive research content, noting that Goldsmith-Pinkham's ability to teach at the right level --- for applied social scientists, not engineers --- is what makes the series exceptional.

The post doubles as a reflection on what makes AI tools hard to teach: demonstrations quickly run over time, audience assumptions are unclear, and much of the work is just "speaking plain English into a terminal." Cunningham notes his own experience presenting to the Federal Reserve Board of Governors and running 90 minutes on a 60-minute slot.

  • Relevance to Economics Research

This is a curated pointer to what Cunningham considers the best video series on Claude Code for empirical economists. It also offers a useful typology of the skills that make researchers effective with AI agents: domain expertise ("taste"), econometric sophistication, and programming fluency. The discussion of the "centaur" archetype --- researchers who combine all three --- is relevant to thinking about who benefits most from AI tools and how the profession is evolving.