Proof review
Skill MattArtzAnthro/AI-Anthropology-Toolkit/skills/proof-review
Use this skill when a publisher returns a typeset proof that has to be checked against the manuscript that was submitted. Triggers include: "my proofs arrived," "check the page proofs," "compare the PDF proof to my Word file," "galley proofs," "the publisher sent proofs and I have five days," "did the copyeditor change my meaning," "answer the author queries," "what corrections should I send the publisher," "is this proof safe to approve." Covers comparing proof against manuscript word by word, separating typesetting artifacts from real discrepancies, auditing citations and the reference list, inspecting rendered pages for production faults, checking that pseudonyms, quoted speech, and orthography survived, and producing a classified discrepancy table plus a correction list in proof-form language. Do NOT use for comments and tracked changes in a returned .docx (use manuscript-markup), for a reviewer report or rebuttal (use academic-review), or for drafting prose (use research-writing).From its SKILL.md
npx -y skills add MattArtzAnthro/AI-Anthropology-Toolkit --skill proof-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
SKILL.md
16.8 KB, as published. Nobody here has run it
Proof Review
Audit a publisher's typeset proof against the manuscript that was submitted, and produce a correction list that can be sent as it stands.
Proof stage is the last point at which an error is still the author's to catch and the first point at which it becomes permanent. It is also the stage most often done by reading, which does not work: reading finds sentences that sound wrong and misses the digit that changed in a citation year, the participant's grammar a copyeditor tidied, and the diacritic that did not survive font embedding. What finds those is comparison.
The comparison is mechanical and this skill will run all of it. What stays with the author is what a difference means.
This skill adopts the Friction by Design conventions at Tier 2; the canonical form is defined in skills/DESIGN.md.
Quick Reference
| Task | Reference |
|---|---|
| The governing comparison rule, the dual-channel method, the full scope inventory, the literal word-and-punctuation checklist, and the normalization rules that say what is not a discrepancy | Read references/comparison-method.md |
| What production can do to pseudonyms, quoted speech, orthography, italicized vernacular terms, and figures, and why each is an evidentiary rather than a stylistic matter | Read references/ethnographic-integrity.md |
| Auditing every in-text citation against the manuscript and the reference list, and every reference entry field by field | Read references/citation-and-reference-audit.md |
| The six classifications, the executive assessment, the discrepancy table, the correction-list wording, the four closing sections, and the reliability requirements | Read references/proof-report-format.md |
What This Skill Will and Will Not Do
A proof correction is an instruction the author sends over their own name, and approval is a signature. Both are recommended here and neither is made here.
Will not do, under any setting. Approve a proof, or state an approval verdict as a decision rather than a recommendation. Decide that an uncertain change is acceptable house style, because a change the files cannot resolve is reported as unresolved rather than waved through. Decide whether a copyedited passage still says what the author meant, which is a judgment about intent that only the author holds. Decide whether an altered or restored detail breaches an anonymization commitment or a consent agreement; name the consequence and stop. Silently repair either document, or report a page, figure, or table as checked when it could not be read. Send anything to the publisher.
An approval the author cannot defend is worse than a late one, and the publisher's deadline is not a reason to convert an open question into a settled one.
When asked for any of these directly, propose the options and ask which.
Will do, on request. Extract both documents and run the full comparison. Apply the normalization rules. Classify every finding. Assemble the discrepancy table. Draft the correction list in proof-form language. Draft responses to author queries from material the author supplies. Format any deliverable. Anything produced this way is marked in the output, and every report carries three registers: Unresolved, Assembled rather than authored, and Not examined.
The third register is the one that matters most here. A proof audit that does not say what it could not read is claiming a completeness it does not have.
Workflow
Step 1: Establish What Arrived and What Governs It
Collect, as facts rather than as decisions, and in one pass:
- The final manuscript submitted to the publisher, and whether tracked changes in it are to be treated as accepted. Default to accepted, and confirm.
- The proof, and whether it is a first proof, a revised proof, or final pages. A revised proof is checked against the previous correction list as well as the manuscript.
- Any copyedited version the author saw between the two. Where one exists, a change may have been accepted at copyedit rather than introduced at proof, which changes the correction, not just the classification.
- The publisher's correction policy and deadline. Presses typically bill or refuse author alterations beyond correcting errors, and that limit shapes which findings can be sent, not which findings are reported.
- The house style, if stated. An unstated house style is inferred from the proof and labeled as inferred.
- Any queries already in the manuscript, as comments, marginal notes, or a copyeditor's query list. Extract them in full before comparing anything.
That last item is a step of its own, and skipping it is the most consequential omission available at this stage. Queries raised on the manuscript are supposed to be answered by the time the proof exists, and the proof shows only the ones that were not. A query resolved differently from how it was asked, or dropped without being answered, leaves no trace on the page at all. Build the list from the manuscript, check each item against what the proof actually contains, and classify each as answered, answered differently than asked, or silently dropped. The second and third are findings, and the third will not be found any other way.
Step 2: Set Up the Comparison
Read references/comparison-method.md.
Use all three reads. Text extraction gives the exact string comparison. Rendered pages give everything extraction destroys or invents: ligatures, discretionary hyphens, dropped glyphs, figure placement, clipped text. The annotation layer gives the queries and marks that are in the file but not in the page content, including the author's own from an earlier sitting.
No single read is a proof audit, and a finding from extraction alone is confirmed against the rendered page before it enters the report.
Work the full scope inventory in the reference rather than the body text alone. Front matter, contributor biography, acknowledgments, running heads, and the copyright statement are where uncorrected errors live longest, because nobody reads them at proof.
Step 3: The Literal Pass
Compare word for word and character for character, then punctuation and typography, against the checklists in the reference.
Apply the normalization rules before reporting anything. Line wrapping, justification, end-of-line hyphenation, ligature encoding, and quotation-mark typography are typesetting, not discrepancies. The exception is stated in the reference and is load-bearing: report a hyphen or dash change when it alters a compound term, a proper noun, a citation, a number, or an established spelling.
Do not treat house style as automatically correct. Establish for each change whether it is harmless, inconsistent, grammatically questionable, or substantively wrong.
Step 4: The Ethnographic Integrity Pass
Read references/ethnographic-integrity.md.
This is the pass a general proofreading procedure does not contain, and it is where the costly errors in anthropological work actually occur. A copyeditor's consistency sweep is the most dangerous thing that happens to a pseudonymized manuscript, and a compositor's cleanup is the most dangerous thing that happens to a transcribed quotation.
Check pseudonym and place-name integrity, quoted speech and transcription conventions, diacritics and non-Latin orthography, the italicization of vernacular terms, figures and their captions, and the acknowledgments.
Anything in this pass that touches anonymization or consent is escalated, never batched, and never resolved by inference from the files.
Step 5: Citations and References
Read references/citation-and-reference-audit.md.
Audit every in-text citation against both the manuscript and the reference list, then every reference entry field by field. A consistent conversion to the publisher's citation style is not a finding unless it introduces an error or an inconsistency.
Report what each file actually contains. Do not repair either one silently, and keep an error inherited from the manuscript distinct from an error introduced at proof: they get different corrections and one of them is the author's own.
Step 6: Visual Inspection of Rendered Pages
Inspect every rendered page for production faults, per the checklist in references/comparison-method.md.
Unresolved author queries and compositor instructions printed as publication text are the two findings that most often survive to print, because both look like they belong to someone else's job. Crop marks, printer marks, and normal publication furniture are not faults.
Step 7: Classify, Report, Recommend
Read references/proof-report-format.md.
Classify every finding into one of the six categories, produce the executive assessment and the discrepancy table, then the correction list in publisher-facing language, then the four closing sections ending in a stated approval recommendation.
Acceptable house-style changes stay out of the correction list and stay in the report. The author decides what to send; they cannot decide about a change they were never shown.
Step 8: Hand Off
- Comments and tracked changes in a returned document: the manuscript-markup skill. Marks anchored to spans are a different object from a typeset page.
- A reviewer report, a decision letter, or a rebuttal: the academic-review skill.
- Rewriting a passage the copyedit damaged: the research-writing skill.
- Whether a restored detail breaches what participants agreed to: the informed-consent skill.
Parameters
- Genre — journal article, edited-volume chapter, or monograph chapter. Chapter proofs add chapter number, contributor biography, running heads, and figure renumbering against the final chapter number; article proofs add ORCID, keywords, and licensing statements.
- Risk posture — where fieldwork was with vulnerable or surveilled participants, the ethnographic integrity pass runs first rather than fourth, and every finding it produces is critical until the author says otherwise.
- Formality register — a first proof takes the full audit; final pages after corrections take a targeted check of the correction list plus a fresh full pass on any page that was reflowed, because reflow introduces new faults on pages nobody changed.
Guardrails
- Never state that the documents are identical unless every page, paragraph, citation, reference, caption, and supplementary element has actually been checked. Partial audits are reported as partial.
- Never report an extraction artifact as a textual error without confirming it on the rendered page.
- Never infer that a difference is harmless because it resembles house style. Resemblance is not intent, and category 6 exists for this.
- Never silently correct an error that was already in the submitted manuscript. Keep four states distinct: introduced at proof, inherited from the manuscript, an intentional publisher change, and uncertain.
- Never regularize a participant's speech, and never let a regularization pass unreported because it reads better.
- Never resolve an anonymization question from the files. The consent agreement is not in either document.
- Name every limitation explicitly. A page that would not render, a figure too low-resolution to read, a table that extraction mangled: each goes in the Not examined register, and the audit is not called complete.
Common Failure Modes
| Failure mode | Prevention |
|---|---|
| Reading the proof instead of comparing it | Step 3 is a mechanical comparison; reading is what missed the error at every earlier stage |
| Extraction artifacts reported as production errors | Dual-channel method; every extraction finding confirmed on the rendered page |
| Mechanical typesetting conversions reported as discrepancies | Normalization rules applied before the report, not after |
| A real dash or hyphen change buried in the normalization rules | The stated exception: compounds, proper nouns, citations, numbers, spellings |
| House style accepted because it is consistent | Consistency is evidence of intent, not of correctness |
| A copyeditor's tidying of quoted speech passing as a style matter | Its own pass, and its own guardrail |
| Front matter, biography, and running heads never checked | They are in the scope inventory, and they are checked first, not last |
| The correction list padded with house-style changes | Category 5 is reported and excluded from the list by rule |
| Deadline pressure converting an open question into an approval | Named in the division of labour; the recommendation states unresolved items rather than absorbing them |
| An audit called complete when three figures were illegible | The Not examined register, which is a required section rather than a caveat |
| A manuscript query silently dropped instead of answered | Step 1 builds the query list from the manuscript and checks each against the proof; a dropped query leaves no mark on the page |
| The author's own earlier annotations reported back to them as findings | The annotation layer is read first and attributed before the pages are read |
| A URL "corrected" or "broken" on the strength of a line-break hyphen | URLs and DOIs are exempt from de-hyphenation and are resolved visually or not at all |
Examples
Example 1: An edited-volume chapter proof, five-day turnaround
Input: "The press sent proofs of my chapter and I have five days. Can you check them against what I sent?"
Output approach: collect the submitted file, the proof, the correction policy, and whether tracked changes were accepted. Run the literal pass and find 140 differences, of which 118 normalize away as typesetting. Run the ethnographic integrity pass and find that a copyeditor regularized a participant's grammar in two block quotations and that a glottal stop rendered as a straight apostrophe throughout. Audit citations and find one year mismatch and one reference cited but absent from the list. Report six critical findings, nine editorial, three reference, four production, and two requiring the author's confirmation, with a correction list of twenty-four items in proof-form language and a recommendation of safe to approve after the listed corrections.
Example 2: A change that may or may not be an error
Input: "The copyeditor changed 'the women said' to 'the women stated' throughout. Is that wrong?"
Output approach: classify as category 6 rather than as an editorial error. The change is consistent, grammatical, and compatible with the house register, so the files cannot establish whether it is acceptable. State what it does to the text, that reported speech verbs carry a stance toward the speaker and that "stated" reads more formally than what a participant did, and ask whether the author wants it reverted. Do not revert it in the correction list before the author answers, and do not let it pass silently because it is consistent.
Example 3: A proof that is not safe to approve
Input: "Everything looks fine to me, I just want to sign off."
Output approach: run the audit anyway and report that the chapter's figure 3 caption was reattached to figure 4 during renumbering, that a pseudonymized clinic appears under its real name in one author query printed on page 11, and that two pages would not render. Classify the first as production, the second as critical with an anonymization consequence named and routed to informed-consent, and the third into the Not examined register. Recommend not safe to approve until the substantive issues are resolved, and state plainly that the recommendation is a recommendation and the signature is the author's.