Resources — Watch-Along Take-Home

Every external link from the talk, in one place

A consolidated list of every external link from the talk plus a few that didn’t fit on slides. Print, save, or fork; the deck embeds these but this file is your reference card.

Last updated: 2026-04-27

Anthropic canonical (start here)

These are the source of truth. If a claim in the talk seemed surprising, check these first.

  • Best Practices — Short, current, very good. The single most useful page for new users.
  • What’s New — Weekly feature digest. Bookmark; check monthly.
  • Memory / CLAUDE.md — The CLAUDE.md walkthrough source.
  • Skills — Skill design, frontmatter, dynamic content.
  • Hooks Guide — All lifecycle events, handler types.
  • Permission Modes — bypass, auto, plan, dontAsk; protected paths.
  • Sub-agents — Task, context: fork, agent definitions.
  • Plugins — /plugin marketplace, install/uninstall.

The two repos from this talk

Other applied-academic Claude Code resources

Highest priority

  • Paul Goldsmith-Pinkham — Claude Code mini-series at Princeton (March 2026, Markus’ Academy, 5 episodes). Applied-econ angle. Topics: getting started, data analysis (Census), web scraping (SEC EDGAR), large datasets (Parquet/DuckDB), writing & thinking. The most authoritative public material on this topic for our field.
  • Claes Bäckman — Claude Code in VS Code, For Academic Economists — Written practical guide (~3,500 words). Setup → recommended extensions (R / Stata / Python / LaTeX) → file formats → CLAUDE.md and skills → git → performance. Recommended companion to this talk: written-form, IDE-focused, complementary to Goldsmith-Pinkham’s video format.
  • Aniket Panjwani — YouTube channel (ongoing, 2026). Northwestern PhD economist. Free 3-hour course for beginners + practical videos: AI Agents for Economic Research, Should economists use Claude Code, Codex, or Cursor?, plus a Stata-specific guide and a deep-dive Claude Code Full Course (2026) slide deck covering CLI, skills, MCPs, plugins, subagents, and a full app build. Pairs well with Goldsmith-Pinkham — different teaching style, more tool-comparison content, more app-building. Site: aniketpanjwani.com.
  • Mihail Velikov — AI in Business & Economic Research wiki (Penn State Smeal, ongoing). Curated knowledge base, ~147 articles organized into 10 sections: Foundations, Prompt Engineering, AI Agents, Claude Code Skills, Data Analysis, Academic Publishing, Finance & Econometrics, Tool Comparisons, Institutional, Productivity. The best single landing page if you want one curator’s reading list across the whole space.

Forks of claude-code-my-workflow

  • hugosantanna/clo-author — Hugo Sant’Anna’s authoring-side fork. 30+ journal profiles, theorist agents, worker-critic pairs. Source of /checkpoint and /review-paper --peer patterns (used with permission).

Standalone community skills

  • jusi-aalto/strategic-revision — DAG-based peer-review response planning. Different approach than /respond-to-referees (Part 6); worth knowing about as an alternative for complex R&R rounds.
  • jusi-aalto/crossref — Bibliography validation against the Crossref REST API. Returns canonical APA citations + DOIs + match confidence.

Adjacent ecosystem

  • chrisblattman/claudeblattman — Chris Blattman’s guide for non-technical academics. Less code-heavy, more workflow-heavy.
  • scunning1975/MixtapeTools — Scott Cunningham’s TikZ + Rhetoric of Decks materials. Source of our TikZ infrastructure.
  • Andrej Karpathy — LLM Wiki (gist) — Karpathy’s pattern for a persistent LLM-maintained markdown knowledge base. Core insight: instead of RAG-style chunk retrieval (where the LLM rediscovers knowledge on every query), maintain durable wiki pages that accumulate over time. New sources get compiled into pages; cross-references update; answers cite the wiki. Pairs naturally with /pdf-to-md (convert sources first) and the claude-code-my-workflow template’s research skills. Open-source implementations: lucasastorian/llmwiki, Astro-Han/karpathy-llm-wiki, kfchou/wiki-skills (Claude Code skills), NicholasSpisak/second-brain (Obsidian).
  • karpathy/autoresearch — Andrej Karpathy’s constraint-based autonomous-research experiment. Aspirational; different shape than ours.
  • EveryInc/compound-engineering-plugin — Every Inc.’s “Compound Engineering” plugin (works in Claude Code, Codex, Cursor, Copilot). 36 skills + 51 agents organized around an ideate → brainstorm → plan → work loop. A kindred-spirit project to our orchestrator pattern — different framing, similar discipline. Worth browsing if you like ours.

Document preprocessing (PDF → markdown)

Claude reads markdown 50–70% more efficiently than PDFs (token count + signal-to-noise). Convert at intake, feed Claude the .md. Both tools are referenced inline in Part 1’s “How Claude reads input” slides.

  • microsoft/markitdown — Microsoft-maintained Swiss-army knife. Converts PDF, DOCX, XLSX, PPTX, HTML, EPUB, CSV, JSON, XML → markdown. Mature, Python CLI, one-line install. First choice — covers ~90% of cases. pip install markitdown && markitdown paper.pdf > paper.md.
  • opendataloader-project/opendataloader-pdf — PDF specialist. XY-cut++ reading order (handles multi-column papers correctly), bounding boxes per element (citation-ready for RAG), 0.907 benchmark accuracy (top open-source). Use when MarkItDown chokes on a complex layout — multi-column journal articles, dense tables, math-heavy PDFs.
  • /pdf-to-md skill in this repo — wraps both tools with a triage step that auto-falls-back from MarkItDown to OpenDataLoader-PDF when the first pass looks suspect.

Applied curriculum + tutorials

Cross-tool

  • openai/codex-plugin-cc (OpenAI, 2026) — the canonical Codex plugin for Claude Code. README has full command reference, examples, and changelog. Three command families: /codex:review (standard), /codex:adversarial-review (challenges the approach), /codex:rescue (delegate a stuck task). Authoritative source; bookmark it.
  • OpenAI Developer Community announcement — original announcement post by VB Srivastav and the OpenAI team. Useful for the framing, but the GitHub README is more current.
  • codex-setup.md (this repo) — step-by-step install + first-use guide for the Codex plugin in VS Code. Marketplace add → plugin install → /reload-plugins → /codex:setup → !codex login → first adversarial review. Includes the three-modes section and a flag reference. Print or bookmark; this is the practical companion to the Part 7 cross-tool slides.

Replication-policy references (Part 5)

Selected field-specific journal policies

For field-specific journal policies beyond those listed above, refer to the journals’ own author-instruction pages directly — those policies are evolving rapidly.

Methodology references (cited inline in the deck)

  • Dhuliawala, S. et al. (2023). Chain-of-Verification Reduces Hallucination in Large Language Models. arXiv:2309.11495. Source of the CoVe / /verify-claims pattern.
  • Stock, J. & Yogo, M. (2005). Testing for Weak Instruments in Linear IV Regression. In Andrews, D.W.K. (ed.), Identification and Inference for Econometric Models. Source of the F > 10 weak-instrument heuristic referenced in Part 6’s demo.
  • Andrews, I. (2026). Some Thoughts on AI and Research. MIT note, April 3 2026. Source of the “bottleneck has shifted from generating output to evaluating it” framing in Part 1.
  • Gloaguen, T., Mündler, N., Müller, M., Raychev, V., & Vechev, M. (2026). Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents? arXiv:2602.11988 (ETH Zurich + INSAIT). Built AGENTbench — 138 instances across 12 real Python repos. Found that LLM-generated context files reduce success ~3% and add 20%+ to step count + cost, while human-written ones give only ~4% gains, and only when narrowly scoped to non-inferable details. Cited in Part 1’s “Length budget” slide; reinforces our “comment out a line, see if Claude breaks” test as empirically grounded.

Field-level framing on AI agents in economics

These don’t appear in the deck but are the most useful one-stop reads on why agentic AI matters for economics research, written by economists for economists.

  • Korinek, A. (2025). AI Agents for Economic Research. NBER Working Paper 34202. The canonical “where the field is” piece. Documents what agents do well (literature review, code generation, data analysis, workflow orchestration) and what they don’t.
  • Grundl, S. (2026). Claude Code as an Empirical Economist: Like Humans but Without the Tails. SSRN. Compares Claude Code (Opus 4.5) to human economists on causal inference tasks under matched instructions. Reports + code at claude-code-economist.com.
  • Yin, Z. et al. (2024). Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance. arXiv:2402.14531. Empirical micro-result: impolite prompts hurt LLM performance, but extreme politeness doesn’t help — there’s a middle band, and it varies by language.

Tools used in the talk

Three things to do, by priority

If you’re going to do one thing this week:

Write a CLAUDE.md for your next paper. 30 minutes. The whole rest of this stack pays off from there.

If you have a free Saturday morning:

Watch Goldsmith-Pinkham’s first 2 episodes and clone Claude-Mini. Different teaching style, different angle — much faster to internalize the pattern with two takes.

If you have a complex R&R with many interlinked concerns:

Try jusi-aalto/strategic-revision before your next response letter. Different approach than /respond-to-referees — DAG-based dependency mapping that surfaces which concerns block which others.

Reach me

If you adopt this stack and write your own skill / fork or extension, send a link. The community section in claude-code-my-workflow’s README grows from emails.