Claude Code for Academics: APSA 2026 Presentation (Spina)
- Author/Source: Alessandro Spina, University of Technology Sydney (UTS Finance) — 3rd APSA Workshop on Quantitative Methods 2026
- Original: https://www.alessandro-spina.com/claude-code/
Key Ideas¶
- Claude Code = LLM + Tools running inside your project folder: reads files → runs commands → produces output. Framed as a dedicated "RA" rather than a chatbot.
- The Amnesia Problem: Claude Code forgets everything between sessions. Solution: build external memory in markdown files —
CLAUDE.md(ground rules, key decisions),README.md(directory map),session_log.md(what was done and why, one file per session named by date). - The Context Window Problem: Quality degrades as the context fills; use
/compactproactively, run one task per session. - Project scaffold: CLAUDE.md (ground rules), README.md, DATA.md (source, license, AI-upload safety per dataset), SCRIPT REGISTRY.md (one line per script: inputs, outputs, question answered), PINBOARD.md (to-dos, ideas, data issues).
- The Cunningham Conjecture: Cross-language replication (R = Stata = Python to 6 decimal places) as a verification strategy. The DGP for coding errors is orthogonal across languages, so mismatches reveal bugs that single-language review misses.
- Skills are reusable instruction files (markdown); Slash commands are explicitly invoked one-shot actions. Skills say "whenever this topic comes up, apply these conventions"; commands say "do this now."
- Examples of skills built:
/spin-up(brief Claude at session start),/wrap-up(save session log),/code-sweep(audit code vs. paper),/paper-editor(seven-audit editorial review),/data-profiler,/glossary,/pinboard. - Failure modes and defenses: fake citations → local PDFs only; wrong estimator → cross-language verification; sycophancy → adversarial personas and agent teams.
- Safety: scope access to one project folder; CLAUDE.md specifies what Claude can/cannot do; start in Plan mode (never YOLO mode); use settings.json allow/deny rules; version control via Git/Dropbox. Run Claude in a Docker container (Goldsmith-Pinkham's
claude-container) to prevent any file access leakage. - Data privacy: only files Claude actively reads enter the API context. Use Claude Code for code (referencing file paths), not raw data. Document AI-upload status per dataset in DATA.md.
- The bottleneck is you: Claude generates output faster than you can verify. One task at a time; read everything before re-prompting.
- Four habits for effective AI use: (1) Start in Plan mode, (2) Define terms (ask Claude to explain jargon in its own words — "translation errors" not hallucinations), (3) Be specific, (4) Separate action from judgment (AI drafts; you decide what is true).
- ROI framing: use AI when (labor saved + verification cost + privacy risk + maintenance cost) < setup cost. High-ROI: data cleaning, figures, code auditing, slide formatting. Low-ROI: novel theoretical arguments, one-off tasks faster to do by hand.
Summary¶
Spina's APSA 2026 presentation is a comprehensive, practically grounded introduction to Claude Code for academic economists and social scientists. Delivered at the 3rd APSA Workshop on Quantitative Methods, it builds from first principles — what an AI agent actually is, and why it's different from a chatbot — through full project scaffolding, skill construction, safety protocols, and honest assessment of failure modes and bottlenecks.
The presentation is distinguished by its focus on systems thinking rather than one-off prompts: the project scaffold (CLAUDE.md, DATA.md, session logs, script registry, PINBOARD.md) creates external memory that survives the amnesia problem across sessions. The Cunningham Conjecture on cross-language replication offers a practical falsification strategy for AI-generated code. The slide deck was itself built with Claude Code — Spina wrote the bullet points, Claude wrote the LaTeX, iterating together — making it a meta-example of the workflow it describes.
Honest downsides are named directly: the bottleneck is the researcher's attention, not Claude's speed; cognitive switching costs make parallel sessions hard to read carefully; and workflow optimization can become a productivity trap. The final framing — "the mechanical work gets faster; the thinking takes the same time it always did" — is one of the clearest articulations of the human-AI collaboration balance in this wiki.
Relevance to Economics Research¶
This is one of the best single-resource overviews of Claude Code for academic economists, particularly for those in empirical fields (political science, economics). It directly addresses the core concerns of applied researchers: data privacy, replication and verification, avoiding fake citations, maintaining reproducible project structure, and the ROI calculation for adopting AI tools. The session-log and script-registry recommendations are immediately actionable. The Cunningham Conjecture cross-language verification strategy deserves wide adoption.
Related Concepts¶
- concepts/claude-code
- concepts/claude-md-files
- concepts/agentic-workflows
- concepts/ai-limitations
- concepts/data-privacy
- concepts/human-in-the-loop
- concepts/cross-language-replication