econ-writing-skill: an Agent Skill for economics papers

Guidance text packaged as an installable skill

Econ Writing Skill is an open-source Agent Skill, built to the Agent Skills open standard, that gives AI coding assistants background on how to write economics papers. The README describes it as distilling advice from over 50 guides by John Cochrane, Deirdre McCloskey, Jesse Shapiro, Keith Head, Marc Bellemare, Claudia Goldin, Edward Glaeser, Michael Kremer and others into a single structured skill file.

The interesting part is what that means in practice. There is no Python package to import and no API to call. The repository's top level shows the shape of it: skills/, sources/, scripts/, evals/, tests/, examples/, an INSTALL.md, and client specific directories named .claude-plugin/, .claude/, .codex-plugin/ and .agents/. Python is the primary language, but that reflects the tooling around the rules rather than a runtime library you would depend on.

So the unit of value is the content itself. If you already have strong conventions for your field, the marginal gain is smaller. If you are a graduate student or a coauthor outside economics who has to imitate a journal's house style, the explicit section formulas and style rules are the point of the project.

Installation is a prompt that makes the client do the work

There is no package manager step. The documented installation method is to open Codex or Claude Code in any project and paste a prompt:

Install or update Econ Write as a standalone skill for the client I am using. Read and follow https://github.com/hanlulong/econ-writing-skill/blob/main/INSTALL.md. Handle same-client migration, backups, and verification yourself. Do not install for or change the other client. Do not ask me to run commands. Finish with a concise result.

The design decision here is that the assistant, not the user, executes the install. The prompt delegates the mechanics to INSTALL.md, explicitly assigns migration, backups and verification to the agent, forbids touching the other client, and forbids asking the user to run commands. The same prompt is reused for a first install and for updates or migration away from a native plugin.

One method per client is the rule the README states. Because the client directories differ,.claude-plugin/ and.codex-plugin/ are separate installations rather than one shared copy, and the prompt is written so a Codex session never modifies a Claude Code setup. That isolation is a sensible default for anyone who uses both tools on the same machine.

Every major section gets a formula to follow

The structural guidance is the densest part of the skill. Each section of a paper carries its own template, and the README names them: an abstract with a four part formula, an introduction using Head's formula, a model section following Glaeser and Varian, a data section, a conclusion with a three part formula, and a title with its own evaluation criteria.

The instruction is that applied empirical, pure theory, mixed theory-empirical, structural and descriptive papers each get tailored structure rather than one generic outline. Theory papers are supported with model presentation advice, proposition writing and Glaeser's approach of starting with an example.

Field conventions are separated out as well, with macro pointing at calibration tables, impulse response functions and DSGE, trade at gravity models and PPML, development at CONSORT and cost-effectiveness, and finance at Fama-MacBeth and winsorization. For a coauthor who is unsure which conventions a field expects, that mapping is more useful than a general reminder to write well. It is also the part most likely to date, since conventions change and a skill file has to be edited by hand to follow them.

Thirteen identification strategies, each with its own guidance

Identification receives the most granular treatment in the README. Writing advice is provided for 13 strategies: RCT, difference-in-differences including the staggered case, IV, RDD, synthetic control, synthetic DiD, bunching, shift-share and Bartik instruments, event studies, machine learning for causal inference, structural estimation and descriptive work, plus guidance for papers that combine several strategies.

The value is in the checklist quality of it. Each strategy carries specific advice on how to present it, which is where most empirical drafts are weakest: the reader has to be told what the identifying assumption is, why it holds, and where it could fail. A tool that knows the conventions of each design gives you a concrete thing to fix rather than a general reminder to be careful.

The same holds for the standard modern practices listed in the skill: pre-registration, multiple testing, specification robustness, and transparency and reproducibility. Combined with the AEA replication package standards, including a README template, data citation, directory structure and code reproducibility requirements from the Data Editor, the skill covers the reproducibility rules that journals now ask about.

A review mode that simulates three referees and scores you

The skill includes a review and audit mode that produces simulated feedback from three perspectives: a Methodologist, a Field Expert and a Writing Critic. That feedback comes with a 100-point scoring rubric, so the output is a number with a breakdown rather than a list of vague suggestions.

Around that sits a 24-item pre-submission checklist spanning content, style and modern standards, and an evaluation test suite of 18 test cases for benchmarking output quality. The repository also ships examples/before-after.md, which is the sort of artifact that makes the style rules concrete, since it shows the transformation rather than describing it.

The anti-AI writing rules are the most distinctive part of the style guidance. They target banned words, uniform sentence length and generic phrasing, and they draw on McCloskey and Cochrane for the underlying rules: active voice, concrete language, triangular structure and reader-first writing. There is a title evaluation framework that scores titles on clarity, specificity, length and memorability with good and bad examples, an economic significance framework that translates coefficients into units, policy benchmarks and back-of-envelope calculations, and citation integrity guidance for verifying references and distinguishing a working paper from its published version. The README states that the skill works in four modes without enumerating them, so the mode boundaries live in the skill files rather than in the README.

Grants, referee replies, policy briefs and journal conversion

The skill reaches past the paper draft itself. Job market paper guidance covers title, abstract, introduction polish, length and signalling breadth. Grant proposal writing covers NSF and ERC proposals, including feasibility, budget justification, timeline and selling a plan rather than presenting findings. Referee response guidance addresses the point-by-point structure of a revise-and-resubmit.

There is also a working paper to journal conversion section with cutting strategy and journal-specific length norms for AER, AER Insights and REStat, plus appendix organisation, dissertation structure for the three-essay format, survey and review paper conventions for JEL, JEP and handbook chapters, and coauthorship conventions such as voice consistency and a designated voice editor.

Formatting is covered too, with a LaTeX guide for tables using booktabs and threeparttable, figures, bibliographies using natbib, and submission requirements for AER, QJE, Econometrica, REStud and JPE. There is guidance on AI use disclosure under AEA and Econometric Society policies.

Two companion projects sit beside it: econ-paper-review-skill reviews a finished draft like a journal referee and produces a referee report with a revision plan, and econ-slides-skill turns a paper into a LaTeX and Beamer talk with a timed speaker script. Adopting only this skill is reasonable if you write papers and not talks.

No releases, no changelog, and a MIT licence with a manual update step

Maintenance for this project is informal in one important way. The repository has no GitHub releases, so there is no version number to pin and no changelog to read when behaviour changes. The last push was on 2026-07-20. The update path is the same prompt as the install path, with INSTALL.md handling updates, migration from a native plugin and troubleshooting, and the README warns to use only one installation method per client.

That means the upgrade cost is a prompt run, and the risk of an upgrade is that your rules change underneath you without a version note telling you what changed. Pinning a commit is the only lever available.

The licence is MIT, which is permissive for personal or research use and compatible with linking into an existing project, though the underlying guides are written by named authors whose own rights are a separate question the README does not address. The project also links to openecon.ai as its homepage.

For a first use, start with one narrow request rather than the audit mode. Asking the skill to write a conclusion for an RDD paper gives you a small output to judge against your own style, and the audit mode is more useful once you know what a good result from it looks like.

Editorial conclusion

Adopt econ-writing-skill if you write economics papers in Claude Code or Codex and want a repeatable structure, a fixed identification strategy checklist and a referee-style review pass before submission. Skip it if you want a bibliography tool, a LaTeX build pipeline or citation database, since this project contributes guidance text rather than either. Before relying on it, read INSTALL.md to confirm which client it installs into, check the skills/ directory to see the rules actually shipped in your copy, and note that the repository has no GitHub releases and the last push was on 2026-07-20, so pin a commit when reproducibility matters.

Frequently asked questions

Official sources

1. hanlulong/econ-writing-skill on GitHub
2. Issues
3. License: MIT
4. Project website
5. README