Most academics use AI the same way they used Google in 2005 - copy a question, paste the answer, hope for the best. Alessandro Spina, a finance researcher at UTS, thinks that's a waste. His 67-slide presentation, "Claude Code for Academics," lays out a structured system for turning Claude Code into what he calls "a dedicated team of RAs who reads your data, runs code, builds slides, and works with you from start to finish."
The guide isn't theoretical. Spina built the entire presentation deck using the workflows he describes, and he shares specific prompts, file structures, and safety configurations throughout.
The Three-Tier System
Spina organizes adoption into tiers, and his most practical advice is blunt: most academics should stop at Tier 1.
Tier 1 is just a CLAUDE.md file (a plain-text rulebook Claude reads at the start of every session), a session logging command, and a script template command. That alone solves the biggest pain point - Claude forgetting everything between sessions.
Tier 2 adds skills (background knowledge files that teach Claude your methodological preferences) and path-scoped rules (different instructions for different file types - never modify raw data files, always include headers in R Markdown). Spina's advice: build these one at a time, only when you catch yourself repeating the same instructions.
Tier 3 brings in agent teams, adversarial review personas, and cross-project orchestration. He references Pedro Sant'Anna's five-phase production workflow from Emory, which uses 10 specialized agents reviewing in parallel with quality gates scored 0 to 100.
The Cross-Language Verification Trick
The most clever idea in the deck is what Spina calls "The Cunningham Conjecture" - named after economist Scott Cunningham. The argument: if Claude writes your analysis in R and introduces a subtle bug, asking it to replicate the same analysis in Stata or Python will likely produce a different bug. When outputs match across three languages to six decimal places, your confidence that the code is correct goes way up.
This is essentially using AI hallucination patterns as a feature rather than a bug. Each language has different syntax and different failure modes, so agreement across all three becomes a strong verification signal.
The "Editor" Persona
Spina shares a custom persona file that turns Claude into a hostile academic reviewer - part "demanding dissertation advisor," part "Referee 2." It runs a seven-dimension audit covering abstract quality, argumentation logic, prose clarity, citation completeness, and a "So What Test" that checks whether readers can answer five fundamental questions about the paper.
The persona never edits files directly. It writes formal reports to a correspondence folder, and the researcher decides what to accept. Spina runs this in a loop: generate report, plan fixes, implement, re-review. The second review, he notes, "chastises you for not fixing first-round issues."
The Safety Section Is Worth Reading
Spina doesn't skip the risks. He covers data privacy (when Claude reads a CSV, its contents go to Anthropic's API - potentially violating data licenses or IRB agreements), sycophancy (Claude reinforcing your priors because it's optimized to be agreeable), and fake citations (plausible but nonexistent references).
His practical defenses: scope Claude's access to a single project folder, never use YOLO mode, keep personal files on a separate Dropbox account, and consider running Claude inside a Docker container using Paul Goldsmith-Pinkham's claude-container tool.
The presentation and associated files - including the Editor persona template - are available on GitHub. For researchers already paying for Claude Pro or Max and wondering how to move beyond copy-paste prompting, this is the most structured onboarding guide available right now.