| name | revise-applied-paper |
| description | Referee and revise an applied economics paper (.qmd, .md, or .tex) against Bellemare (2020), "How to Write Applied Papers in Economics." Three steps — cold review, written referee report, revision with response memo — run once interactively or for several rounds with --loop N. Use when the user asks to revise, review, referee, polish, or check a paper draft, an introduction, abstract, data section, results section, tables, or conclusion against applied-econ writing norms. |
Revise an applied economics paper (Bellemare 2020)
Simulate a journal R&R: a cold referee reviews the paper against references/bellemare-checklist.md, writes a referee report, and the paper is then revised against that report with a response memo. All artifacts persist next to the paper so rounds can build on each other.
Supporting files (relative to this skill's directory):
references/bellemare-checklist.md — the distilled guidance; the referee must apply every item.references/bellemare-fulltext.txt — full text of Bellemare (2020); consult only when the original wording is needed.references/uw-abstract-guide.md — abstract-writing criteria from the UW–Madison Writing Center; applied alongside §8.3 of the checklist.assets/referee-report-template.md — the report format; fill it exactly, never invent a different structure.assets/response-memo-template.md — the response memo format.assets/BellemareHowToPaperSeptember2020.pdf — source of record; do not read during reviews.
Arguments
/revise-applied-paper [file] [--loop N]
file — the manuscript. If omitted, look in the current directory for .qmd, .md, or .tex files that look like a manuscript (not README, notes, slides). If several candidates exist, ask which.--loop N — run up to N autonomous rounds (see Loop mode). Without it, run one interactive round.
Artifacts
All bookkeeping lives in a referee/ folder next to the manuscript:
referee/
├── referee_report_<YYYY-MM-DD>_r<N>.md # one per round
├── response_memo_<YYYY-MM-DD>_r<N>.md # one per round
├── needs_author.md # ledger: substantive items only the author can fix
└── declined.md # ledger: items the author struck, with reason
Get the date with date +%F. Determine the round number N by counting existing referee_report_* files for this manuscript + 1.
Step 1 — Cold review
Dispatch the review to a fresh subagent (Agent tool, default type) so the referee reads the paper cold, without this conversation's context — real referees haven't watched the author write the draft. Give the subagent ONLY:
- The manuscript path (instruct it to read the whole file; for
.tex, also any \input/\include sub-files; for .qmd/.md, the YAML front matter).
- The checklist path:
references/bellemare-checklist.md — instruct it to walk every item, section by section, and record PASS (cite where) / FAIL (cite the offending passage or gap) / N-A (say why) for each.
- The style sweeps to run across the whole text: past tense for data/results (present tense is the norm outside the conclusion); passive voice; bland enumeration of descriptive means for mere controls; Stata/R code names ("AGE_2", "edu") in tables or prose; too many decimals, scientific notation, unintuitive units; "robust standard errors" without saying robust to what, clustering without level and rationale; causal language anywhere the identification strategy does not support it.
- Rounds ≥ 2: the paths of all prior referee reports, response memos,
needs_author.md, and declined.md, plus these standing rules — verify whether each prior comment was resolved; do not re-litigate a resolved comment unless the fix is wrong; do not re-raise anything in declined.md or needs_author.md; rounds ≥ 2 focus on major comments — no new comma-level minor comments after round 1 unless a revision introduced them.
Have the subagent return findings as a structured list: per finding — checklist item, verdict, severity (major/minor), evidence with location, suggested action, fixable-by-editing (yes/no), and for rounds ≥ 2 a verification verdict per prior comment.
Step 2 — Referee report
Fill assets/referee-report-template.md with the subagent's findings and save it as referee/referee_report_<date>_r<N>.md. Severity rules:
- Major (M1, M2, …): identification gaps (the three endogeneity sources, SUTVA), missing balance tests / robustness / heterogeneity / mechanisms / limitations, missing data provenance, bait-and-switch or formula-violating introduction, overclaiming, structural problems.
- Minor (m1, m2, …): tense, voice, notation, table formatting, decimals, variable names, roadmap, abstract polish — anything fixable by editing alone.
Interactive mode: STOP here. Show the user the report (or its summary plus the comment list) and ask which items to strike. Record struck items in referee/declined.md with the user's reason. Only surviving items proceed to Step 3.
Loop mode: do not stop. The standing rule replaces the checkpoint — see Loop mode below.
Step 3 — Revise
Work through the surviving report comments one by one, in report order:
- Apply directly (Edit) every comment marked fixable by editing: tense and voice; section order and names; rewriting the introduction into Head's formula order using only material already in the paper (add a roadmap if missing); tightening the abstract toward hook/question/value-added; table titles, plain-English variable names, significance-symbol notes, decimal consistency (in the source — flag if tables are generated by code elsewhere); equation notation (Greek coefficients, Latin variables, subscripts, no re-used estimand notation). Title changes: propose 2–3 options in the memo; change it only if the user picks one.
- Park every comment that needs new analysis, new data, or the author's judgment: append it to
referee/needs_author.md (comment id, what is needed, why). Never fabricate content to "fix" these — no invented citations, numbers, results, or robustness checks.
- Preserve the source format:
.qmd/.md as Markdown, .tex as LaTeX; never break code chunks, cross-references, or citations.
Then fill assets/response-memo-template.md mapping every comment to applied / parked / declined, and save it as referee/response_memo_<date>_r<N>.md.
Loop mode (--loop N)
Run rounds autonomously: Step 1 → Step 2 → Step 3, repeat. Defaults: cap at N rounds (suggest 3 if the user gives no number).
- Checkpoint replacement: auto-apply only fixable-by-editing comments; everything substantive is parked in
needs_author.md untouched. Nothing is ever auto-declined — declined.md is written only by the user.
- Stop early when a round produces no new major comments (parked and previously-known items don't count as new).
- Anti-churn: every round's referee gets the full prior paper trail and the no-relitigation rules from Step 1.4. If a round proposes reverting a change made in a previous round, do not apply it — flag the conflict in the memo instead.
- Finish with a consolidated summary in chat: rounds run and why the loop stopped; everything applied across all rounds (grouped by paper section); the full
needs_author.md list as the user's to-do; any flagged conflicts.
Final report (both modes)
End with: where the report, memo, and ledgers live; what changed in the manuscript (by section); what's parked for the author as concrete next actions; and the three acid-test verdicts from the latest report.
