manuscript-review

Files (1)

• SKILL.md

SKILL.md (first lines)

# manuscript-review

This file is the canonical Empirical Workflow Kit implementation. Runtime views
defined in `workflow.manifest.yaml` link here; edit only the canonical tree.

Resolve `<skills_root>` from `workflow.manifest.yaml:canonical_source.skills_root`
before using a path below.

Adapted from `ericluo04/claude-academic-workflow` at commit `8958cc246e65cdf7c36604f397a1c1719b7e2c14`; see `THIRD_PARTY_NOTICES.md`.

Six agents read the manuscript in parallel and return severity-tagged findings with quoted
evidence; the main thread consolidates them into one dated report ranked by what would sink the
paper. Nothing in the manuscript is edited. Topology and agent roles adapted from
[`claesbackman/AI-research-feedback`](https://github.com/claesbackman/AI-research-feedback), the
reader-understanding assessment from
[`aspi6246/Claude-Code-Presentation`](https://github.com/aspi6246/Claude-Code-Presentation).

For research organization, economic meaning, contribution, and behavioral claims,
read `../empirical-workflow/references/research-writing.md` before synthesis or
drafting. Apply the relevant evidence-strategy boundary; do not turn the
operation's output format into a mandatory manuscript structure.

## Options

| Argument | Default | Meaning |
|---|---|---|
| a journal name (first token) | `top-field` | journal persona and bar for agent 6 |
| a file path | auto-detect | main `.tex`, or a `.pdf` of the compiled draft |
| `--quick` | off | run only agents 6 and 3, skip the rest |
| `--referee` | off | journal referee mode: a submittable report, see below |

Recognized journals, case-insensitive, covering the UTD 24 and FT 50 lists plus the top economics
and IO journals; full names, publishers, and list membership are in
`<skills_root>/literature-review/references/journal-scope.md`. Information systems: `MISQ`,
`ISR`, `JMIS`, `IJOC`. Operations: `OR` (or `OpsRes`), `MSOM`, `POM`, `JOM`. Marketing: `MKSCI`,
`JMR`, `JM`, `JCR`, `JCP`, `JAMS`. Management and organization: `AMJ`, `AMR`, `ASQ`, `OrgSci`,
`SMJ`, `JIBS`, `JMgmt`, `JMS`, `OrgStud`, `OBHDP`, `JAP`, `ResPol`, `JBV`, `ETP`, `SEJ`.
Current FT 50 additions: `AMA` (Academy of Management Annals), `ASR` (American Sociological
Review), `PsychSci` (Psychological Science). Former FT 50 titles remain recognized as valid target
aliases but are not labeled current FT 50 unless the central scope reference says so.
Cross-field: `MS`. Accounting: `TAR`, `JAR`, `JAE`, `AOS`, `CAR`, `RAST`. Finance: `JF`, `JFE`,
`RFS`, `JFQA`, `RoF`. Economics: `AER`, `AERI`, `QJE`, `JPE`, `Econometrica`, `REStud`, `REStat`,
`JEEA`, `AEJApplied`, `AEJPolicy`, `AEJMicro`, `AEJMacro`, `JME`, `RED`. Industrial
organization: `RAND`, `JIE`, `IJIO`, `JEMS`, `QME`. Anything unrecognized is treated as a file
path, and an absent journal means `top-field`: high general standards for a leading field
journal, no specific persona. Add journals by editing this list. Store the result as
`TARGET_JOURNAL`.

## Referee mode

Before reviewing a third party's submission, complete the protocol's policy and confidentiality
check. If the journal, institution, agreement, or consent terms do not authorize transfer to the
active model or service, stop. Record the authority and permitted scope in the review request.

`--referee` switches the skill from a pre-submission check on the user's own draft to a
submittable report on a paper the user is refereeing for a journal. Enter this mode when the flag
is passed, when the user says they are refereeing, or when the paper is clearly not theirs. When
it is ambiguous whose paper it is, ask which hat the user is wearing before doing anything else.

What changes:

- The manuscript must be supplied as a path or a PDF. Phase 1 uses step 1 only; never glob the
  working directory or the Overleaf projects. If no path was given, ask for one.
- Agent 6 keeps its full brief, including the reader-understanding assessment, but critiques the paper
  for the editor and drops every piece of advice addressed to the author: no fallback outlets in
  part 5, no coaching on what would reach the target's bar. Its recommendation becomes one of the
  journal's conventional decisions: accept, minor revision, major revision, or reject.
- All six agents still run (two with `--quick`), and the CRITICAL/MAJOR/MINOR triage still
  happens, as internal analysis. The final report is written from that analysis and never shows
  the tags.

The output replaces the phase 3 structure. Write `REFEREE_REPORT_<YYYY-MM-DD>.md` to the current
working directory, or wherever the user says; never into an Overleaf project. Address it to the
editor and the authors in the conventional register:

1. A summary paragraph restating the paper in the referee's own words: the research question, the
   design, the data, and the main findings. Specific enough to show the paper was read.
2. Major comments, numbered. Built from the CRITICAL and MAJOR findings: identification threats,
   missing analyses, overclaiming, internal contradictions. Each states the issue, points at the
   exact location, and where possible says what would resolve it.
3. Minor comments, numbered. Notation, exposition, table and figure documentation, typos worth
   the authors' time. Compress agent 1's copy edits into a few representative items; a referee
   report is not a proofreading pass.
4. Recommendation, with a short justification tied to the major comments.

Report back in chat: the report path, the recommendation, and the counts of major and minor
comments.

## Phase 1: find the manuscript

1. An explicit path argument wins. Route PDFs through the helper configured in
   `runtime-profile.yaml`:
   `$EWF_PDF_HELPER text paper.pdf` for the text, or
   `$EWF_PDF_HELPER png paper.pdf --pages N --dpi 150 --out /tmp/p` and then Read
   the PNG when the layout or a figure matters.
2. No path: Glob `**/*.tex` in the working directory.
3. Still nothing: search the manuscript roots configured in `runtime-profile.yaml`, show the
   matching project directories, and ask which one. Do not guess between projects.

In referee mode, step 1 is the only step: skip steps 2 and 3 and ask for the path if none was
given.

When parallel workers are unavailable or not authorized, execute the same six lenses sequentially
into separate immutable findings files. Do not let later lenses read earlier verdicts. Consolidate
only after all requested lenses finish; consequence, not vote count, sets priority.

Then build the file set. The main document is the `.tex` containing `\documentclass` or
`\begin{document}`; follow every `\input`, `\include`, and `\subfile` to get the rest, and read
all of them. Figures are `**/[Ff]igure*/**/*.pdf` and the same for `png`, `jpg`, `jpeg`, `eps`,
`svg`, plus root-level images. Tables are `**/[Tt]able*/**/*.tex` plus root-level
`*[Tt]able*.tex`. Exclude `**/_minted-*/**`, `**/build/**`, `**/output/**`, `**/.git/**`, and
Dropbox conflicted copies (`*conflicted copy*`), which are stale and produce phantom findings.

Record every file path and its role, the figure and table lists, and the title, authors, and
abstract. If no figures turn up, tell the user they may live in a non-standard directory and to
re-run with an explicit path. Same for tables, adding that agent 5 will then be limited to
captions and cross-references.

## Manuscript-only cold read

From bobbyZhong10/empirical-workflow-kit, MIT.

More from bobbyZhong10/empirical-workflow-kit