proofread

Review the paper and write findings to report.md. The text itself stays untouched until the user asks to apply items.

PK means python3 ${CLAUDE_PLUGIN_ROOT}/scripts/paperkit.py, run from the project root.

Two branches by argument:

• apply <IDs> (e.g. apply R3 R7, apply all high): go to Apply.
• anything else: a scope (a .tex file, or a section name you resolve to its file and lines), or nothing for the whole paper. Go to Review.

Review

1. Load the rules. Load the writing skill and read its pattern list ${CLAUDE_PLUGIN_ROOT}/skills/writing/ai-patterns.md, plus ## Style and ## Terminology from CLAUDE.md.
2. Run the scripts. PK sync, then PK check [files] --json. The check output is a list of candidates with suggested severity and category.
3. Read report.md. Note the highest ID used. Items marked done or wontfix are settled: do not report them again. Open items whose location lies in the scope are replaced by this run; open items outside the scope stay.
4. Read the scope in reading order (main file, then \input order) and review every paragraph, caption, footnote and table for:
   • Content: unsupported or overstated claims, gaps in the argument, results in the text that differ from the tables, missing baselines or details a reviewer would ask for.
   • Citations: each \cite against the paper it cites. Use the literature skill: the claim must match the notes or the paper body. Papers not read-full get read now. A citation that no body passage supports is 🔴. Also: claims that need a citation and have none.
   • Language: spelling, grammar, articles, agreement, the English variant from paper.config.json.
   • AI-Style: the patterns of the writing skill.
   • Consistency: terminology, abbreviations, notation, capitalization, tense, numbers.
   • ACL-Compliance: only what is visible while reading (e.g. no Limitations section; author names in a review version of an acl-arr paper, but not for venue semeval, where papers are not anonymous; check paper.config.json). The full check is /acl-paper:guidelines-check. Where a finding depends on intent you cannot read from the text (is this term deliberately different? is this the claimed contribution?), collect the questions and ask them with AskUserQuestion in one round before writing the items.
5. Confirm the script candidates in context. Drop false positives (an abbreviation that is a model name, a quote inside \texttt). Several identical low findings become one item listing all places.

6. Write the items into report.md in the format its header comment defines, under the severity heading, with new IDs counting up from the highest used. Each item quotes the text and proposes a concrete fix.

   Severity: 🔴 high misstates something, breaks a rule that blocks submission, or a citation that its source does not support. 🟠 medium is what a reviewer would flag. 🟢 low is polish.

Done when every paragraph in scope was read, every citation in scope checked against its source, and every script candidate confirmed or dropped. Then report in chat: counts per severity and category, and the three items to fix first.

Apply

For each requested ID (all high = every open 🔴 item):

1. Find the location. Lines may have moved: locate the quoted text.
2. Make the smallest edit that resolves the issue, following the writing skill. Citation items follow the literature skill.
3. Set the item's status to done and move it under ## Closed.

When an item has several reasonable fixes, or needs the authors' decision (content, results, formatting), ask with AskUserQuestion before editing. When it turns out wrong, leave it open, add a line - **Note:** ..., and say so. Report the applied edits as before → after.