科研 Skill
文稿校对润色
proofread
Proofreads a legal draft’s prose for typos, grammar, and internal consistency (hyphenation, defined terms, cross-references). Use for a proofing pass; not a cite-checker or editor.
SKILL.md
Proofread
An internal proofreading pass over an article draft — the prose itself, not the citations. Deterministic scripts do extraction, chunking, and the mechanical lint (doubled words, footnote-call-before-punctuation, unbalanced quotes/parens, supra/infra note-number direction breaks, Part-label existence, heading-outline rules) for near-zero tokens; model agents read every sentence chunk-by-chunk, two whole-document agents adjudicate consistency (hyphenation, spelling, capitalization, terminology drift, TOC staleness, roadmap-vs- structure), and a mandatory skeptical re-verification of every candidate finding keeps false positives out of the report (the cite-check lesson: no finding ships unvetted).
Scope of the citation side (recalibrated 2026-07-14 after the Codex-memo comparisons):
the deterministic layer DOES check mechanical citation typography and internal citation
consistency — unfilled _/__ placeholders, Id.., §-spacing, U. S., signal commas,
mixed signals ("see also generally"), mid-cite (citing Id.) capitalization, missing
terminal punctuation, page-range dash style, pincites below a cite's own first page,
year-shaped pincites ("596 U.S. 845, 2017 (2022)"), full case cites missing their decision
year, U.S.-Reports volume↔year sanity (585 U.S. can't be 2013), the same case cited with
different volumes across footnotes (signal words stripped before keying — the Jardines
lesson), Name, supra note N targets that don't contain that name, and a ≥12-word passage
repeated nearly verbatim in a nearby paragraph (duplicated-quotation residue). None of that
needs an external source.
SHORT FORMS (added 2026-08-10 on reader feedback): a full citation repeated within
Bluebook R. 10.9's five-footnote window where the short form belongs; a case short form
stranded past that window, or with no full cite before it anywhere; a secondary source
cited in full twice when R. 4.2 calls for Author, supra note N. All decidable from the
draft alone — see ../legal-legal-research/bluebook-typeface-and-short-forms.md
for the rules as applied.
TYPEFACE (added 2026-08-10, .docx only): the extractor now resolves run properties
(w:i, w:smallCaps, w:caps) through character and paragraph styles, so italics and
small caps ARE checked — signals/Id./supra left roman, the note N after an italicized
supra wrongly italicized, italics spilling onto the comma before a citation, small caps
broken mid-word, and case-name italicization against R. 2.1(a) (roman in a full cite
standing in a citation clause, italic in a textual sentence or short form). This also
fixes the reverse problem: an ALL-CAPS Part header is stored as ordinary text with the caps
coming from the style, so proofers used to argue about the capitalization of words the
reader never sees that way. Those paragraphs now carry a display flag and are off-limits
for capitalization findings.
What it is NOT: citation/quotation verification against real sources (that's
/cite-check — the two compose: proofread the prose, cite-check the authorities), a full
Bluebook linter (signal order, citation order, and full-cite composition stay out of
scope), or a substantive editor. Suspicious text INSIDE quotations is never asserted as a
draft error — it ships as an author query ("verify this quote against its source").
Indentation and block-quote layout remain invisible; on the markdown fallback path the
whole typeface layer is switched off rather than guessed at.
How to run
1. Parse the request
- draftPath — the draft (required).
.docxis canonical;.md/.txtwith pandoc footnotes works. - deep — TWO independent proof passes instead of one: extract twice (1800- and
1200-word chunks — different boundaries, so different eyes on every passage) and pass the
concatenated chunk lists as
chunkFiles; fuzzy dedup + verify absorb the overlap. Calibration (2026-07-14, NOTES.md) showed single-pass proofreading is stochastic — ~75–80% overlap between two same-model passes, each catching real errors the other missed — so a second offset pass buys far more recall than smaller chunks alone. Also smaller verify batches and a higher soft cap. Turn on for a near-final draft. - style — opt-in: also surface genuine style problems (confusing sentences, unintended repetition), clearly separated in the report. Default OFF — proofreading, not editing.
- fable — opt-in: run the per-chunk proofers on Fable instead of Opus ("call in fable", "use the big model", typically alongside deep for a final pre-submission read).
Model map: proofers Opus (Fable with the flag); consistency, verify, and report always inherit the session model — verify is the false-positive gate and the report is the deliverable, so those never run on a smaller model.
1b. Ask up front — depth and style — in ONE dialog (author's call 2026-08-09; style added 2026-10-27, ../DESIGN-upfront-dialog.md)
Unless the request already settles both, ASK before doing anything else — one
AskUserQuestion carrying up to two questions, up front, so the choices are made before the
snapshot rather than surfaced as an offer after a standard run has already finished. Depth changes the
extraction (step 3 runs the extractor once or twice), so it cannot be changed later
without re-running.
Skip a question the request already settles, by dropping it from the call — never ask it with a preselected default. "Deep", "thorough", "near-final", "pre-submission", "call in fable", "use the big model", "quick pass", "standard" settle Depth; "style", "flag confusing sentences", "editing too" settle Style on, and "just proofread" or "typos only" settle it off. Don't re-ask what they just told you. Also skip on an explicit re-run of a configuration already chosen this session. When both are settled, make no call.
Question 1 — Depth (header Depth), three options in escalating order:
| Option | Args | What to say |
|---|---|---|
| Standard — one pass | deep:false, fable:false | One 1800-word-chunk pass, Opus proofers. Good for a working draft or a quick check. |
| Deep — two offset passes | deep:true, fable:false | Two independent passes (1800 + 1200 chunk boundaries). In one calibration run this surfaced roughly a quarter more confirmed findings at no wall-clock cost (the wider fan-out parallelizes) — one draft's numbers, not a guarantee. For a near-final draft. |
| Deep + Fable | deep:true, fable:true | Deep, plus the per-chunk proofers on Fable. The final pre-submission read. |
Give the real reason in the option descriptions, not just labels: single-pass proofreading is stochastic (~75–80% overlap between two same-model passes — each misses real errors the other catches), so depth buys recall that a better per-pass model does not. Don't tag a default "(Recommended)" — the right tier depends on how near-final the draft is, which the user knows and you don't. If they say, note it and move on.
Question 2 — Style (header Style), in the same call:
| Option | Args | Description shown |
|---|---|---|
| Off (default) | style:false | Proofreading only: typos, grammar, missing or doubled words, consistency. |
| On | style:true | Also flags confusing sentences and unintended repetition, in a clearly separated section. This is editing, not proofreading, and it is opinion; turn it on when you want a second reader's reactions. |
An earlier version of this step kept style out of the dialog because a second decision
"taxes every run." That reasoning predated multi-question dialogs: one AskUserQuestion call
now holds up to four questions in a single interaction, so the tax is one more row in a dialog
that already appears, not a second interruption — and a user who does not know the option
exists never finds it otherwise. Fable stays inside the Depth ladder, not a question of its own.
When AskUserQuestion is unavailable (a headless or scheduled run), use the defaults —
Standard, style off — and state them in the deliver message and the report footer.
2. Snapshot the draft (always)
Live OneDrive/Word co-authoring changes the file between reads — the smoke-test draft gained/lost footnotes across three reads in one morning. Copy the draft to the session scratchpad first and run everything off the snapshot; record the snapshot time for the report header.
cp "<draftPath>" "<scratch>/snapshot.docx" && chmod a-w "<scratch>/snapshot.docx"
Record the snapshot time; the extractor computes the snapshot's SHA-256 (in proof.json)
— both go into the workflow args and appear in the report header as provenance.
The draft is read-only to this skill — structurally, not just by convention. The
original draft's path goes into the cp above and NOWHERE else: never into Workflow args,
never into an agent prompt. Everything downstream reads only the write-protected snapshot's
extraction artifacts, and every workflow agent runs as the Explore type (no Edit/Write
tools). Explore agents do keep a shell, which the workflow uses to grep the snapshot, so
the guarantee rests on the path never reaching an agent and on the snapshot being
write-protected, not on the agent type alone. The ONLY file this skill ever creates outside the scratchpad is the
report (step 5) — a new .md next to the draft, written by you, never the .docx itself.
3. Extract + lint (deterministic — run yourself, near-zero tokens)
python3 ../proofread/scripts/extract_proof.py "<scratch>/snapshot.docx" "<scratch>/proof"
Default chunking is 1800 words — sized so each proofer reads carefully rather than skims.
For deep, run the extractor a SECOND time into a separate outdir with
--chunk-words 1200 and concatenate both runs' chunk_files into the workflow's
chunkFiles (use the 1800 run's proof.json/fulltext.txt as outDir; the snapshot is
identical so paragraph ids line up). Prints a summary (paragraph/footnote/
word counts, chunk_count, lint_finding_count…). It writes <scratch>/proof/proof.json,
fulltext.txt, and chunks/chunk-NN.json.
- If the summary shows
tracked_insertions/tracked_deletionstrue: the extractor reads the accepted text and skips tracked deletions — tell the user and ask whether to accept-all first if the draft is heavily marked. - The summary also reports
formatting_available,typeface_findings,short_form_findingsandall_caps_headings.formatting_available: falsemeans a markdown/plain-text draft — say so at deliver time, since the typeface section will be absent. - Pull the workflow args from
proof.json(compact, no indent):python3 -c "import json;d=json.load(open('<scratch>/proof/proof.json'));print(json.dumps({'chunkFiles':d['chunk_files'],'lintFindings':d['lint_findings'],'headings':d['headings'],'noteParas':d['note_para'],'sha256':d['sha256'],'formatting':d['formatting'],'counts':{'paragraphs':d['paragraph_count'],'footnotes':d['footnote_count'],'words':d['word_count']}}))"
4. Launch the workflow
Workflow({
scriptPath: "../proofread/proofread.workflow.js",
args: {
skillsDir: "..", // where the bundled helper scripts live
outDir: "<scratch>/proof", // agents read fulltext.txt + proof.json here
draftName: "<draft filename without extension>",
chunkFiles: <proof.json chunk_files>,
lintFindings:<proof.json lint_findings>,
counts: { paragraphs, footnotes, words },
headings: <proof.json headings>, // [{para,label,text}] — report location paths
noteParas: <proof.json note_para>, // note -> referencing ¶ (for paths on n.N)
sha256: <proof.json sha256>,
formatting: <proof.json formatting>, // typeface layer; available:false on .md drafts
snapshotTime:"<when the snapshot was taken>",
deep: false,
style: false,
fable: false // true => proofers on Fable instead of Opus
}
})
Pipeline: the whole-document consistency agents launch first and run alongside the
per-chunk proof fan-out — usage, structure, and (on .docx) typeface, the one stage that
reads fulltext_fmt.txt and is allowed to reason about italics; every candidate finding
(proof + consistency + typeface + deterministic lint)
is deduped and then re-derived by a skeptical verify agent that Greps the flagged passage
out of fulltext.txt and re-reads it in context; only confirmed findings reach the report,
sorted into draft order. Spend-guarded (soft cap ~250k lean / ~400k deep); runs in the
background and notifies on completion. No web access needed — the whole run is local.
5. Deliver
The workflow returns { report, counts, exitState, degradation, rejectedFindings }.
- Write
reportto<draft's directory>/<draftName> — Proofread.md(next to the draft). - Research memo header & production footer (required, per the author — global rule in
~/.claude/CLAUDE.md): the report opens directly under the title (alongside the
snapshot line) with the structured header block (
Topic:/Date:/By:); theProduction details:block goes at the very end of the file (after everything else, set off by a---rule) — never up top.Dateis local time (date "+%Y-%m-%d %H:%M %Z"when writing the file);ByisClaude-<orchestrating session model>(exact id);Production detailsnames the skill, the workflow-agent model(s) actually used (read them off the run's per-agent progress entries — don't assume the model map), the wall-clock duration, and the token count — from the workflow completion notification's<usage>block (duration_ms,subagent_tokens,agent_count), labeled subagent output tokens (orchestrator overhead not included). Example header:
Example footer (last lines of the file):Topic: Proofread of <draft name> Date: 2026-07-14 15:05 PDT By: Claude-fable-5--- Production details: Claude Code /proofread; 8m 25s; 664,041 subagent output tokens across 24 agents (workflow agents Claude Opus 4.8) - TL;DR the counts: errors / judgment calls / conventions, and note how many candidates the verify pass rejected (false-positive control working, not something to hide).
- If the source draft's mtime changed since the snapshot, say so — findings cite the snapshot, and paragraph numbers may have drifted.
- Surface
exitStateand anydegradation(e.g. likely-severity findings shipped unvetted under budget pressure) and offer a targeted re-run. - Offer a .docx copy via
python3 ../legal-legal-research/md2docx.py "<report>.md" "<report>.docx" "<Title>". - Note once what was and wasn't covered: on a
.docx, italics/small caps WERE checked against the document's real run formatting (block-quote layout and indentation were not); on a markdown draft, no formatting was visible at all. Citation substance is never checked here —/cite-checkcovers the authorities.
Guardrails (never relax)
- The draft is never edited. Read-only by construction: the original path is used once
(the snapshot
cp) and never reaches the workflow; the snapshot ischmod a-w; all agents run as theExploretype, which has no file-editing tools (it does have a shell, so the first two protections are the ones that matter). This skill reports findings — it never applies them. If the user wants fixes applied, that is a separate, explicit request outside this skill, made against the live draft with their eyes on it. - Quoted language is never a draft error. Archaic spelling and odd grammar in quotes stand, no [sic] suggestions. Transcription-smell inside a quote (garbled grammar, stray characters, duplicated clauses) ships only as an author query — "verify against the source" — never as a fix-the-draft finding.
- Citation SUBSTANCE is out of scope; citation FORM is deterministic-or-typeface only. Footnote prose (parenthetical wording, textual sentences) is proofread; mechanical typography, internal cite consistency, and short-form placement come from the lint layer; italics/small caps come from the typeface stage; signal order, citation order, and whether a cite is correctly composed belong to the journal and /cite-check. Model proofers still skip citation strings except obvious typos.
- Formatting claims come from the formatting layer or not at all. Only the typeface
stage and the verify agents adjudicating its findings may reason about italics or small
caps, and only against
fulltext_fmt.txt. Every other stage sees plain text and must not guess. Indentation and block-quote layout are invisible to all of them. Ifformatting.availableis false (markdown draft), no typeface claim may ship at all. - All-caps headings are never a capitalization finding. Word applies caps via the style, so the stored text of a Part header is unreliable — its casing is not what prints.
- Every finding is re-verified before it ships. A verify agent re-reads the full passage and rejects what a careful editor wouldn't mark; unverifiable findings ship only under an explicit "Unvetted" label, never silently.
- Verbatim find-strings. Every reported finding carries text copied character-for- character from the draft so the author can Cmd+F straight to it; if the string can't be found verbatim, the finding doesn't ship.
- No internal ¶ numbers in the report (author, 2026-08-23). The extractor's paragraph ids
exist only for the agents (grep anchors, dedup, sort keys) — the author editing in Word
cannot see them, so they locate nothing. Reader-facing locations are the containing
Part/Section, an approximate percent-through-the-document, and visible footnote numbers
("note N"); the verbatim find-string is the primary locator. The workflow's
locLabel()renders these — keep any future report changes on that convention. - Proofread, don't edit. No style/wordiness/clarity opinions unless
stylewas requested, and then clearly separated as optional.
Prerequisites
Dispatches via Claude Code's Workflow tool and a bundled proofread.workflow.js
script (step "Launch the workflow"). In an agent without an equivalent tool, run the
same stages (extract/chunk/lint deterministically, then the two reading passes)
sequentially instead of expecting the Workflow call to succeed.
Limitations
"Cmd+F" (the delivered find-string) is Mac phrasing; the same string works with Ctrl+F on Windows/Linux — the report's point is the verbatim string, not the key combination.
示例提示词
帮我校对这篇文章草稿的语法、错别字和内部引用一致性。
所属插件
安装该插件会一并获得同领域的 13 个 Skill / MCP。
同插件的其他科研 Skill
- Zotero文献库构建Builds or extends a Zotero collection for a project: finds sources, adds metadata via the Zotero Web API, and runs an anti-hallucination check. Use to populate a Zotero library with verified sources.
- 裁判文书实证研究Guides empirical research on a sample of Chinese court judgments: sampling frame, coding scheme, inter-coder reliability, and descriptive statistics on rulings. Use for a quantitative case-law study.
- 最高法院案件材料整理Gathers the full document set for a US Supreme Court case and loads it onto the user’s reMarkable in a per-case folder. Use when prepping a SCOTUS case for a podcast, class, or article.
- 文稿引注核查Verifies every citation/quotation in a legal draft against CourtListener, official reporters, and Zotero/the open web, returning a graded Bluebook footnote report. Use to cite-check a US law draft.
- 代码专利特征扫描Scans a codebase for distinctive, potentially patentable design patterns with structured scoring and evidence. Use before a patent consultation, not as legal advice.
- 代码专利检索验证Turns code-scan findings into patent-database search queries for researching existing implementations before consulting an attorney. Not legal advice.
- 快速引注查找Finds 2-4 verified US case-law or scholarly citations with a pincite for a single proposition, footnote-ready with Bluebook form. Use for a quick cite-finding request, not a full memo.
- 中国法律全领域检索China-focused systematic legal research across statutes, regulations, judicial interpretations, and guiding cases, tailored by user role. Use to judge legality or find a legal basis/damages standard.
- 深度法律研究备忘录Fans out multiple agents across free US legal sources, verifies every authority against CourtListener/Justia/Google Scholar, and returns a polished memo. Use for US case-law research.
- 专利构思特征分析Analyzes a described concept to surface what makes it distinctive, with structured scoring for a patent consultation. Not legal advice.
- 专利构思检索验证Turns a concept analysis into patent-database search queries for researching the prior-art landscape before consulting an attorney. Not legal advice.
- 法律研究插件初始化配置One-time, conversational setup for the legal-research plugin: saves and tests a CourtListener token and checks Zotero credential availability. Use to set up this plugin’s CourtListener access.
常见问题
文稿校对润色能做什么?
Proofreads a legal draft’s prose for typos, grammar, and internal consistency (hyphenation, defined terms, cross-references). Use for a proofing pass; not a cite-checker or editor。
如何安装文稿校对润色?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「文稿校对润色」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。
文稿校对润色需要哪些权限?
该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。
文稿校对润色经过官方审核吗?
是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。