Writing with AI

The role of AI in academic writing

AI writes fast and clean. It does not write with insight. Here is how I divide the labor:

• You supply the argument, the interpretation, the economic intuition.
• The AI handles structure, formatting, style enforcement, and mechanical editing.

This chapter covers how to use AI for manuscript work without producing prose that reads like it was generated by a machine.

The writing standard

The template follows McCloskey’s Economical Writing principles. These are baked into CLAUDE.md, so the AI follows them automatically:

1. Active verbs. “Prices increased” — not “an increase in prices occurred.”
2. Concrete language. “Machines and workers” — not “capital and labor inputs.”
3. Plain words. “Use” — not “utilize.”
4. Delete ruthlessly. Cut “very,” “basically,” “actually,” “it is important to note that.”
5. One word, one meaning. Pick a term and keep it throughout the paper.
6. End strong. The last word in a sentence carries the emphasis.
7. No boilerplate. Never open with “This paper discusses…” or “The remainder of this paper is organized as follows.”
8. Tables and graphs are writing. Same rules apply to labels, titles, and notes.

AEA style

The template includes manuscript/aea_style_guide.md with detailed AEA formatting rules. The AI reads this file before making any edits to the manuscript. Key points:

• Equations: numbered, referenced as “equation 1” (lowercase, not abbreviated)
• Tables: all in the body (not an appendix), landscape if needed, with descriptive notes
• Figures: high-resolution, clear labels, source notes
• Citations: author-date format via natbib
• Numbers: spelled out below 10 in text, digits in tables

How to use AI for writing tasks

Drafting sections

Give the AI your argument and let it draft:

> Write the data section. We use CDC WONDER mortality data
> from 2015-2022, county-level, restricted to accidental drug
> poisoning deaths (ICD-10 X40-X44). We merge with ACS county
> demographics. Explain the sample construction and any
> exclusions. Keep it under 400 words.

Review the draft. The AI will follow your style guide, but you need to verify that the substance is correct. Add your interpretation. Cut anything generic.

Editing existing prose

> Edit section 3.2. Tighten the language. Apply McCloskey
> rules. Don't change the argument, just the writing.

The AI will shorten sentences, activate passive voice, remove filler words, and improve flow. Review every change — occasionally the AI will alter meaning while trying to improve style.

Tables and figures

> Write the LaTeX for Table 1. It should show summary statistics
> for treatment and control counties: mean, sd, min, max for
> population, income, death rate. Add a note explaining the
> sample and time period.

The AI writes the .tex fragment and saves it to output/tables/. The manuscript \inputs it directly.

Citation management

If your Zotero auto-export is configured, the AI can add citations:

> Cite Ruhm (2019) for the claim about economic conditions
> and drug mortality. Use the BibTeX key from references.bib.

The AI searches your .bib file, finds the key, and inserts the \cite{} command.

What to watch for

• Bland hedging. AI defaults to “may,” “could,” “potentially.” Replace with direct claims supported by your evidence.
• False precision. AI sometimes invents specific numbers or citations. Verify every factual claim.
• Generic framing. Opening paragraphs that could describe any paper. Delete and write your own.
• Over-signposting. “In this section, we will discuss…” — just discuss it.

A useful test — Read the AI’s draft aloud. If it sounds like a committee wrote it, rewrite it. Good academic writing has a voice. The AI can help you find yours by cleaning up your rough drafts, but it should not replace your voice entirely.

Auditing your draft: Referee 2

Once you have a working draft — organized, coded, with tables and figures in place — you need to stress-test it before submitting. You cannot grade your own homework, and the AI that helped you write it should not be the one reviewing it.

Referee 2 is an open-source audit protocol created by Scott Cunningham (of Causal Inference: The Mixtape) that uses AI to perform five systematic audits on your research. For Cunningham’s own account of how he uses Claude Code for empirical research and where Referee 2 fits, see his Claude Code for Empirical Research post.

The five audits:

1. Code audit. Identifies coding errors, missing value problems, and logic gaps in your scripts.
2. Cross-language replication. Creates independent implementations of your analysis in R, Stata, and Python. The key insight: hallucination errors are likely orthogonal across AI-produced code in different languages. If your Stata script has a subtle bug, an independent R implementation will likely produce a different bug — making discrepancies easy to catch by comparing outputs to six or more decimal places. You cannot find errors by re-reading the same code that produced them.
3. Directory audit. Checks that your replication package is submission-ready: every file accounted for, every path working, every dependency documented.
4. Output automation audit. Verifies that all tables and figures are programmatically generated from code, not manually edited.
5. Econometrics audit. Reviews specification coherence, identification strategy, standard error choices, and sample definitions.

The key principle is independence. Referee 2 operates as a separate session — it reads your project but never modifies your code. It produces a formal referee report and saves replication scripts to code/replication/ for your review.

How to use it

After your draft is complete and your pipeline runs cleanly:

1. Open a new AI session in your project folder.
2. Load the Referee 2 persona from the MixtapeTools personas directory.
3. Point it at your manuscript, scripts, and output.
4. Review the report. Fix what needs fixing. Repeat.

Think of it as a pre-submission peer review — rigorous, structured, and designed to catch the things you are too close to the project to see.