Barrios Skills — John Barrios's curated workflow for economists and accountants. Prioritize reproducible empirical work, clear identification language, and journal-ready output.
Econ / finance / accounting pre-submission referee
John Barrios's referee workflow for this collection. Evidence rules and specialist lenses are adapted from Ingar30/reviewer (MIT). This skill stays a single referee workflow. It does not run that repository's parser, validators, or multi-agent runtime.
After the judgment is settled, use econ-write + econ-humanizer to implement revisions.
When to use
- Stress-test a working paper before journal submission
- Produce a referee-style report the author can act on
- Separate confirmed corrections (the paper disagrees with itself) from qualifications and optional development
Skip this skill for a summary, a rewrite, or proofreading alone.
Evidence contract
Every major comment cites manuscript evidence. Do not invent table numbers, coefficients, sample sizes, signs, or formulas.
- If the PDF, table image, or equation is too garbled to check, mark Cannot verify and name the missing exhibit. Do not reconstruct it.
- One independently fixable defect per comment. Distinct denominators, samples, or corrections stay separate even when they share a sentence.
- State the problem in the first line. Example: "Table 3, column 2 reports 0.041; the text on p. 12 calls the same cell 4.1 percent."
- A missing row, undocumented covariate, or absent code file is a disclosure limit. It is not proof the authors omitted the procedure. Ask for the exhibit unless retained evidence shows the error.
- Before calling two quantities inconsistent, confirm they share estimand, aggregation, sample, denominator, and period. A ratio of averages need not equal an average of ratios. Percent and percentage-point changes can both be correct. A log-point coefficient is not automatically an arithmetic-mean effect.
- An explicitly scoped comparison need not show every arm, outcome, or subgroup. Flag an exclusion only when it misleads the stated estimand or departs from a stated plan.
- Separate the observation from the diagnosis. A verified inconsistency may have several causes. Ask the authors to reconcile it. Do not assert an unobserved coding mistake or a replacement number.
- The absence of a method, dataset, experiment, or formal model is not a defect. Skip that lens.
- Literature critiques name the study and the version checked, or they are Cannot verify. An older abstract does not prove a later draft lacks a design.
- Do not accuse fraud. Ask for clarification.
Detail for each lens: references/audit-lenses.md.
Read the manuscript safely
When the input is a PDF, parse it with paper-pdf before the audits:
bashpython skills/writing-and-review/paper-pdf/scripts/parse_paper_pdf.py \ --pdf paper.pdf --out paper-pdf-work
Read parsed/full_text.md, the table and figure crops, and parsed/quality.json.
- If a page or table is marked unsafe, conflicting, or OCR-recommended, use the page image. A blank cell is blank. Do not type a number the parser did not extract.
- Note exhibits you cannot trust. Those claims stay Cannot verify.
- A heading or table missing from extracted text is not evidence the manuscript omitted it. Search the page images before reporting a broken cross-reference.
Workflow
Finish each coverage pass before writing findings. Do not stop at the first error. Stop the whole review only when the exhibits behind the main claim cannot be read.
1. Classify
Empirical reduced-form, archival accounting, asset pricing, banking, macro, theory, structural, experiment, or mixed. Name the claimed identification and the target venue if the user gave one.
2. Contribution sentence
Write the contribution in the author's words, then in skeptical words. If they diverge, that is a comment only when a specific sentence overclaims. Positioning preferences that do not make a sentence false go to exploratory suggestions.
3. Always-on audits
Run these on every paper that reports results. Record findings before later passes. A later interpretive pass must not rewrite a number already checked.
- Source fidelity. Ledger of central claims against the cited table, figure, column, sample, sign, and definition. Then one pass over each main display.
- Numbers. Recompute material differences, percents, percentage points, ratios, log points, and adding-up identities (shares, flows, probabilities, budget items) when the paper says the parts exhaust a total.
- Claims and summaries. Abstract, introduction, and conclusion versus the body, including construct labels broader than the measured object.
- Displayed specs. Dependent variable, sample, treatment, reference group, fixed effects, and estimator versus the prose and table note.
- Cross-references that a result depends on (Table 4, Appendix B, equation (3)).
4. Conditional audits
Run a lens only when the paper contains that object. See the reference file.
Identification, sample construction, robustness language, economic magnitude, power and multiple testing, randomization, institutional or regulatory facts, pre-analysis or replication claims, literature positioning, formal theory, external validity and policy language.
5. Rank, then write
Classify each finding into exactly one tier:
| Tier | What belongs here |
|---|---|
| Confirmed correction | Arithmetic inconsistency, internal contradiction, mislabeled sample or outcome, inaccurate description of the paper's own evidence, undisclosed departure from a stated plan, or a missing definition required to read a central result |
| Material qualification | Design, measurement, inference, or external-validity limit that narrows a claim and depends partly on judgment |
| Exploratory suggestion | Alternative framing, extra literature, optional analysis, or presentation preference |
Rank by tier, then by consequence for the central result, then by how readily the author can act. A subtle identification worry does not outrank a wrong number in the main table. If reasonable readers could disagree about the preferred framing, it is not a confirmed correction.
Prefer the smallest fix. Precise language beats a demand for new data or a new design when language resolves the issue.
Comment format
text[Tier: Correction|Qualification|Exploratory] [Topic: Source|Numbers|Identification|Inference|Data|Magnitude|Contribution|Clarity] Location: §X / Table Y / Figure Z (p. N) Claim checked: ... Evidence: what the manuscript shows, with the compared cells Why it matters: consequence for the central result Minimum fix: smallest adequate correction or qualification Optional stronger check: only if it adds material value; label it optional
If unverified: Status: Cannot verify plus what is established, what is unavailable, and the smallest disclosure that would settle it.
Deliverables
Always produce all three.
A. Referee report
Write for the authors. Precise, calm, constructive. No process notes about agents or tools.
- Summary. One paragraph: contribution, then up to five confirmed corrections, then one sentence on qualifications.
- Confirmed corrections. Three to six when the evidence supports that many; fewer when it does not. Each uses Evidence, Why it matters, Minimum fix. Add Optional stronger check only when useful.
- Material qualifications.
- Exploratory development. Label these as opportunities, not required corrections.
- Cannot verify.
| Item | Established | Unavailable | Next step |
|---|
- Advisory forecast, labeled private pre-submission advice (Accept / R&R / Reject as a forecast). It must follow the tiers. Exploratory comments alone do not support a reject forecast. Confirmed errors that change the main result do.
B. Editing notes
Separate from substance. Copyediting, notation nits, and bibliography maintenance (year suffixes, incomplete fields) live here. A citation problem stays in the referee report only when it changes the identity of a result the paper relies on.
C. Revision plan
Order: fixes possible from existing evidence, then qualifications and documentation, then optional analyses.
| Priority | Comment | Action | Depends on |
|---|---|---|---|
| P0 | … | … | … |
Field flags
| Field | Recurring issues |
|---|---|
| Archival accounting | Accruals and disclosure constructs broader than the measure; Compustat filters and fiscal-year alignment; auditor or office clustering; human-coded versus model-coded text; adding-up of earnings components |
| Corporate finance | Endogenous policy; bad controls; industry×year fixed effects; silent clustering switches; leverage and governance exclusion stories |
| Asset pricing | Multiple testing; tradability; microstructure; anomaly t-stats versus economic magnitude in return space |
| Banking | Regulatory timing; call-report breaks; call versus market data; coverage dates in the data dictionary |
| Governance / ESG | Construct validity; boilerplate text measures; selection into disclosure |
Venue checklists
Use when the user names a target outlet. These are pre-submission stress tests.
JAR / TAR / JAE
- The question is about accounting measurement, disclosure, assurance, or real effects of reporting rules
- Constructs are defined against the obvious alternative measures
- The identification threat is named in accounting terms (discretion, concurrent standards, auditor change, mandate timing)
- Compustat
indfmt/datafmt/popsrc/consol, fiscal-year alignment, and winsorization are disclosed where the sample depends on them - Clustering matches where the shock lives (firm, auditor or office, state)
- Magnitude is in accounting units, not only a t-stat
- Mechanism evidence is present or the claim is narrowed
JF / JFE / RFS
- One sentence states the contribution against the relevant frontier
- The exclusion story is explicit
- Fixed effects and clustering match the variation and do not switch silently across tables
- Pricing papers address multiple testing, tradability, and microstructure, or scope them out
- Magnitudes are in return space with the horizon stated
- Staggered adoption is not only two-way fixed effects when timing varies
AEA / general interest
- External validity and mechanism get weight alongside the estimate
- The identification section can stand alone for a non-specialist empirical reader
- Main tables carry the claim; robustness forests sit in the appendix
If the paper straddles fields, say which bar you are applying and why.
Red lines
- Do not rubber-stamp a weak design
- Do not rewrite the paper into a different paper
- Do not demand citations you cannot name accurately
- Do not put parser or extraction failures on the author
- Do not let agreement with your own earlier impression replace a check against the cited cell
