version: "1.0.4" name: proofread description: "Use when you need academic proofreading of a LaTeX paper (11 check categories)." allowed-tools: Read, Glob, Grep argument-hint: [project-path or tex-file]

Academic Proofreading

Report-only skill. Never edit source files — produce reviews/<scope>/proofread/<YYYY-MM-DD-HHMM>.md only (where <scope> is the paper slug, e.g. paper-jtp).

Output Path

Per rules/review-artefact-routing.md (auto-loads in research projects (path-scoped to paper-*/ and paper/)):

• Source slug: proofread
• Write reports to: reviews/<scope>/proofread/YYYY-MM-DD.md inside the project, where <scope> is the paper slug (e.g., paper-jtp). Path is relative to the research project root, not the Task-Management repo.
• Never at project root (./CRITIC-REPORT.md-style filenames are forbidden — pre-rule layout).
• Idempotency: if today's file exists, append a same-day descriptor ({date}-revision.md, {date}-r2.md, {date}-pre-submission.md) — never overwrite.
• Index update: if reviews/INDEX.md exists, write a one-line entry under "Latest per source" pointing at the new file. Otherwise /review-recap will rebuild the index next time it runs.
• Infrastructure repos (Task-Management, atlas-workspace, etc.): this section does not apply — the path-scoped rule won't load there.

When to Use

• Before sending a draft to supervisors
• Before submission to a journal/conference
• After major revisions to check consistency
• When you want a fresh-eyes check on writing quality

When NOT to Use

• Formal audits — use the Referee 2 agent for systematic verification
• Argument quality — use /devils-advocate for logical scrutiny
• Citation completeness — use /bib-validate for bibliography cross-referencing (though this skill flags obvious citation format issues)

Workflow

1. Locate files: Find all .tex files in the project (and .log files for LaTeX diagnostics)
2. Read the document: Read all .tex source files in order
3. Run 11 check categories (below)
4. Produce report: Write reviews/<scope>/proofread/<YYYY-MM-DD-HHMM>.md under the project directory (where <scope> is the paper slug, e.g., paper-jtp; create the directory if it does not exist: mkdir -p reviews/<scope>/proofread/). Do NOT overwrite previous reports — each review is timestamped to the minute. Canonical convention: ~/Task-Management/docs/reference/review-state-schema.md.

Check Categories

1. Grammar & Spelling

• Spelling errors (including technical terms)
• Subject-verb agreement
• Sentence fragments or run-ons
• Misused words (e.g., "effect" vs "affect", "which" vs "that")
• American English is the default for all papers and conference articles. Flag any British English spellings (e.g., -ise, -our, -re, analyse, scepticism).

2. Notation Consistency

• Variable notation used consistently throughout (e.g., always $x_i$ or always $x_{i}$, not both)
• Subscript/superscript style (e.g., $\beta_1$ vs $\beta_{OLS}$)
• Matrix/vector formatting conventions (bold, uppercase, etc.)
• Consistent use of \mathbb, \mathcal, \mathbf for sets, operators, vectors
• Equation numbering: all referenced equations numbered, unreferenced ones unnumbered

3. Citation Format

• Consistent use of \citet (textual) vs \citep (parenthetical)
• No raw \cite{} when \citet/\citep is available
• Author name spelling matches between text mentions and citation keys
• Multiple citations in chronological or alphabetical order (check which convention)
• No "As shown by (Author, Year)" — should be "As shown by \citet{key}"

4. Academic Tone

• No informal contractions (don't, can't, won't → do not, cannot, will not)
• No first-person overuse (some "we" is fine; excessive "I think" is not)
• No casual hedging ("pretty much", "kind of", "a lot")
• Appropriate use of hedging language ("suggests" vs "proves")
• No exclamation marks in body text
• Consistent tense (present for established facts, past for specific studies)
• Filler hedge phrases to flag (delete on sight): "interestingly", "importantly", "it is worth noting", "it should be noted that", "needless to say", "of course", "obviously", "clearly" (unless something is literally clear from a figure)
• Vague quantifiers to flag: "very", "quite", "rather", "somewhat", "fairly" — replace with specific magnitudes or delete
• Tautological constructions: "significant and important", "novel and new", "unique and distinctive" — one adjective suffices

5. LaTeX-Specific Issues

• Overfull hbox: Check .log file for Overfull \hbox warnings — report line numbers and severity (badness)
• Equation overflow: Long equations that exceed column/page width
• Float placement: Check for [h!] or [H] overuse; prefer [tbp]
• Missing labels: Figures/tables/equations referenced but without \label{}
• Orphan/widow lines: Check for \\ abuse that creates bad page breaks
• Unresolved references: ?? in output indicating broken \ref{} or \cite{}

6. Citation Voice Balance

Check the ratio of in-line (\citet) to parenthetical (\citep) citations:

• Count in-line vs parenthetical citations across the full document
• Flag if ratio exceeds 1:1 (in-line should be the minority) — Major
• Flag runs of 3+ consecutive in-line citations in a paragraph or section — Major
• Flag paragraphs that open with an in-line citation when the author's identity isn't the point — Minor
• Flag "As shown by \citet{}" patterns where parenthetical would be more natural — Minor
• Report the overall ratio (e.g., "42 parenthetical, 28 in-line — ratio 1.5:1")

See docs/reference/conventions.md § Citation Voice Balance for the full convention.

7. TikZ Diagram Review

If the document contains TikZ code (\begin{tikzpicture} or \tikz):

• Label positioning: Labels not overlapping nodes or edges
• Geometric accuracy: Coordinates and angles consistent with intended layout
• Visual semantics: Arrow styles match meaning (solid = direct, dashed = indirect, etc.)
• Spacing: Nodes not too cramped or too spread out
• Consistency: Style matches across all diagrams in the document
• Standalone compilability: Each diagram should compile independently

For detailed spatial verification (Bezier depth calculations, gap minimums, shape boundary clearance), see `../shared/tikz-rules.md`.

8. Numeric Text↔Table Cross-Check

Cross-check every number mentioned in the prose against the corresponding table or figure. Flag ANY discrepancy, no matter how small.

• Coefficient claims: "a 3.2 percentage point increase" in text vs coefficient of 0.031 in the table — flag the mismatch
• Sample sizes: "N = 1,247" in text vs N row in table — must match exactly
• Significance claims: "statistically significant at the 5% level" in text vs the actual stars/p-value in the table
• Summary statistics: means, medians, standard deviations mentioned in prose must match the descriptive statistics table
• Figure references: claims about trends or magnitudes must be consistent with what the figure shows
• Table/figure existence: every table/figure referenced in text must exist; every table/figure in the document must be referenced in text
• N consistency: the number of observations should be consistent across related specifications unless there's an explained reason for differences

9. Causal Language Audit

Audit causal claims against the stated identification strategy. The strength of causal language must match the strength of the research design.

• Flag "X causes Y" or "the causal effect of X" when the identification strategy is observational (OLS with controls, correlational) — should be "is associated with" or "predicts"
• Flag "proves" — almost never appropriate; use "provides evidence consistent with"
• Match language to design:
• DiD/RDD/IV → "we estimate the causal effect" is acceptable with qualifiers
• OLS with controls → "we find a relationship" or "our estimates suggest"
• Correlational → never use causal language
• Flag "significant" ambiguity — must always be clear whether "statistically significant" or "economically meaningful/large"
• Flag coefficient descriptions without units or scale — "a coefficient of 0.031" means nothing without knowing the units of X and Y
• Flag correlation/causation conflation — any statement that slides from an association to a causal claim without an identification argument

10. Equation Completeness

Verify that mathematical notation is complete and internally consistent.

• Every variable in an equation must be defined in the text (either before or immediately after the equation)
• Error terms: properly specified (ε_it vs u_i vs e_ij) and consistent with the econometric framework described
• Equation numbering: sequential, and all cross-references to equations are correct
• Subscript/index structure must match the level of observation — e.g., if the text describes individual-level variation, the equation should have individual subscripts, not county-level
• Summation/expectation indices: verify bounds match the described sample
• Consistent notation across equations: don't switch between β and b for the same coefficient, or between X_i and x_i

11a. Anonymity (double-blind venues only)

If the project's vault submission frontmatter or CLAUDE.md indicates a double-blind venue (CCS, NDSS, S&P, USENIX, ICML, NeurIPS, ICLR, FAccT, AAAI, etc.), run paper-side checks P1–P8 from ~/.claude/skills/_shared/double-blind-anonymity-checklist.md and flag every FAIL as Critical:

• P1 title page anonymized (no \author{} with real names)
• P2 no \thanks{}, \acknowledgements, funding, or grant references in body
• P3 body uses third-person self-reference (no "we previously showed", "in our prior work")
• P4 self-citations are cited in the third person with the real bib entry kept — do not flag a named self-cite entry as a violation; naming the authors in a third-person citation is correct (anonymity comes from the author block, not the bibliography). Blind the entry only if the venue's CFP explicitly requires it (rare; some security venues). See rules/double-blind-self-citation.md.
• P5 self-references use third-person voice: flag first-person ("we previously showed", "in our prior work [N]") near a self-cite — not the author names themselves ("Burnat and [Collaborator] [N] show X" is fine)
• P6 no identifying URLs (personal websites, GitHub repos with handles)
• P7 PDF metadata clean (pdfinfo shows no Author / identifying Subject)
• P8 figures/screenshots have no identifying watermarks

Skip this section only when the user explicitly says single-blind / non-blind.

11. Preprint Staleness

For every citation that looks like a preprint or working paper, check whether a peer-reviewed version has since been published. Flag stale preprints as Major issues.

• Detection signals: URL contains arxiv.org, ssrn.com, nber.org; journal field says "Working Paper", "mimeo"; entry type is @techreport or @unpublished
• Action: note the stale citation and suggest the published venue/year
• This is a lighter version of `/bib-validate`'s preprint check — only flag obvious cases visible from the .bib or \bibitem entries. For thorough preprint checking, recommend running /bib-validate separately.

Severity Levels

Quality Scoring

Apply numeric quality scoring using the shared framework and skill-specific rubric:

• Framework: `../shared/quality-scoring.md` — severity tiers, thresholds, verdict rules
• Rubric: `references/quality-rubric.md` — issue-to-deduction mappings for this skill

Start at 100, deduct per issue found, apply verdict. Insert the Score Block into the report after the summary table.

Recurring Pattern Grouping

When the same issue appears 3+ times, group it as a single pattern finding instead of listing each instance separately. This prevents reports bloated with 50 individual items when the real message is "you have 3 recurring problems."

Format:

### M3: Hedge phrase "interestingly" (8 instances)
- **Category:** Academic tone
- **Locations:** lines 42, 67, 103, 145, 189, 203, 267, 301
- **Problem:** Filler hedge phrase adds no content
- **Fix:** Delete all 8 instances

One deduction for the pattern (not 8 separate deductions). Escalation still applies: 5+ instances of the same minor issue → one Major deduction.

Report Format

# Proofread Report
 
**Document:** [filename]
**Date:** YYYY-MM-DD
**Pages:** [approximate]
 
## Summary
 
| Category | Critical | Major | Minor |
|----------|----------|-------|-------|
| Grammar & spelling | 0 | 0 | 0 |
| Notation consistency | 0 | 0 | 0 |
| Citation format | 0 | 0 | 0 |
| Academic tone | 0 | 0 | 0 |
| LaTeX-specific | 0 | 0 | 0 |
| Citation voice balance | 0 | 0 | 0 |
| TikZ diagrams | 0 | 0 | 0 |
| Numeric cross-check | 0 | 0 | 0 |
| Causal language | 0 | 0 | 0 |
| Equation completeness | 0 | 0 | 0 |
| Preprint staleness | 0 | 0 | 0 |
| **Total** | **0** | **0** | **0** |
 
## Critical Issues
 
[List each with file, line/section, and specific issue]
 
## Major Issues
 
[List each with file, line/section, and specific issue]
 
## Minor Issues
 
[List each with file, line/section, and specific issue]
 
## Quality Score
 
| Metric | Value |
|--------|-------|
| **Score** | XX / 100 |
| **Verdict** | Ship / Ship with notes / Revise / Revise (major) / Blocked |
 
### Deductions
 
| # | Issue | Tier | Deduction | Category |
|---|-------|------|-----------|----------|
| 1 | [description] | [tier] | -X | [category] |
| | **Total deductions** | | **-XX** | |
 
## Recommendations
 
[Optional: overall observations about the writing — prioritise fixes by deduction size]

Council Mode (Optional)

For high-stakes pre-submission checks, run proofreading in council mode to get independent assessments from multiple LLM providers. Council mode surfaces formatting issues that any single model might miss.

Trigger: "Council proofread my paper" or "thorough proofread"

How it works:

1. The main session reads the document and constructs the proofreading prompt
2. The prompt is sent to 3 different models via council-cli (or council-api for API mode)
3. Each model independently runs the 7 check categories
4. Cross-review identifies agreements and disputes
5. Chairman synthesis produces a single PROOFREAD-REPORT.md with council notes

Invocation (CLI backend — free with existing subscriptions):

cd packages/council-cli
uv run python -m council_cli \
    --prompt-file /tmp/proofread-prompt.txt \
    --context-file /tmp/paper-content.txt \
    --output-md /tmp/proofread-council.md \
    --chairman claude \
    --timeout 180

See skills/shared/council-protocol.md for the full orchestration protocol.

Value: Diminishing returns for pure formatting — council mode is most valuable when combined with citation voice balance and notation consistency checks, where different models have genuinely different pattern recognition.

Log to REVIEW-STATE.md (final step)

After writing the proofread report, append a row to the project's REVIEW-STATE.md:

bash ~/.claude/skills/_shared/review-state-log.sh \
  --check proofread \
  --paper "<paper-{venue} dir>" \
  --verdict "<PASS|ISSUES FOUND>" \
  --open-issues "<total-issues-across-categories>/<total-issues-across-categories>" \
  --report "reviews/<scope>/proofread/<YYYY-MM-DD-HHMM>.md" \
  --notes "<one-line: e.g. '3 critical, 12 minor; mostly notation §3'>" \
  [--trigger "pre-submission-report|review-cluster"]

• Verdict: PASS if no issues found across any category; ISSUES FOUND otherwise.
• Open issues: total count across all 7 (or 11) check categories at run time.
• Trigger: pass orchestrator name only if invoked via /pre-submission-report or /review-cluster. Otherwise omit.

Schema: ~/Task-Management/docs/reference/review-state-schema.md.

Cross-References

• `/bib-validate` — For thorough bibliography cross-referencing
• `/latex` — For compilation and error resolution (run before proofreading to ensure the document compiles cleanly)
• Referee 2 agent — For formal code + paper auditing
• `/devils-advocate` — For argument quality and logical scrutiny