科研 Skill
学术手稿章节撰写
drafting-manuscript
Drafts a chapter or article section by section from synthesized wiki content, agreeing each section’s argument first. Use to turn stable research synthesis into prose.
SKILL.md
Drafting a Manuscript
Turn stable synthesis pages into publishable prose. Every claim gets a citation. Every citation resolves in output/bibtex/references.bib. No inventing, no paraphrasing from memory — the wiki is the single source of truth for what is claimed.
But the wiki is deliberately terse — bullets, one-sentence claims, page numbers in parentheses. It is a pointer to the depth, not the depth itself. Drafting straight from it produces dense, compressed prose that reads like reflowed bullet points, without the examples and explanations a reader needs. The fix is not to write more from memory (that reintroduces hallucination) but to reach back to the sources for depth: the illustrative examples, the explanation of why a claim holds, and the surrounding argument live in the source pages' quote/example sections and — when those are too thin — in the original PDFs in the shared library (<library>/pdf/<bibkey>.pdf, put there by acquire-sources). Elaboration must be grounded and cited; only connective/expository framing (transitions, restating an argument's logic) is uncited. See Writing with depth.
And the wiki tells you what is claimed — never what this chapter argues. The claim chain is the author's contribution, not the wiki's — which is why the thesis comes from the author first (STOP 0), and you sharpen it rather than propose it. A thesis the user merely approves looks like agreement but skips the thinking: a polished machine formulation reads as better than one's own and is accepted unchecked (automation bias). Formulating the thesis is the author's work; checking it against the wiki, naming its gaps and its strongest counter-position is yours. And the claim chain is the thing that drifts: drafted in one pass, a chapter wanders into whatever the sources happen to be rich about, and the drift only becomes visible once thousands of words exist and are expensive to throw away. So the argument is settled on paper and discussed before any prose exists, then carried out one section at a time, each section's argument agreed before it is written. See The three stages.
Announce at start: "Using drafting-manuscript to draft <chapter/section> into <path>. Three stages: you state the thesis in your own words, we settle the argument architecture on it, then go section by section — argument, then prose, with a check-in at each."
If a condition is unmet: tell the user which (e.g. "no stable synthesis yet"),
ask for a short reason to draft anyway, write it into
knowledge/_meta/gate-overrides.log, and start the draft. Repeated overrides
on (3) are a maintenance signal.
On (5): an open flag means a content review found an unresolved concern
(overstatement, weak support, stale claim) on a page you are about to turn into
prose — drafting from it would bake the problem into the manuscript. Prefer
resolving it first (fix the page, set the flag state: resolved) over
overriding. Note which page and flag kind in the override reason.
The three stages
Stage B repeats per section. A chapter is finished when every section in the outline has Status: approved.
Moved to references/the-three-stages.md to keep this file short; read it when this part is needed.
When to use
- A plan Draft task is current (
executing-research-planroutes here) - Synthesis page(s) are stable and user asks for chapter/article draft
- Re-drafting a section from scratch after peer review (the outline is the thing you revise against)
NOT for: working through an existing text paragraph by paragraph (use revising-manuscript — it keeps the decisions of that discussion in a ledger), first-draft brainstorming (use brainstorming-research), unsynthesized material (go back and synthesize first), grant research narratives from scratch (use grant-finder).
Checklist
Full steps for each part are in references/checklist/; open the one you need:
- Stage A — Architecture —
references/checklist/01-stage-a-architecture.md: 1. Confirm plan task — find the exact entry in<slug>-plan.md; confirm output file path 2. Pre-flight checks (SOFT-GATE) — stable synthesis pages? BibTeX complete? wiki-lint green? No openreview_flagson the pages you pull fro… - Stage B — Section loop (repeat per section) —
references/checklist/02-stage-b-section-loop-repeat-per-section.md: 9. Sketch the section's argument — claim, steps with their evidence, the concrete material, the counter-position and how it is handled, the handoff to the next section (see Section argument sketch). Write it… - Stage C — Finishing —
references/checklist/03-stage-c-finishing.md: 17. Chapter coherence check — once every section is approved, read the chapter as a whole, because each section passing on its own does not make the chapter cohere (LLM prose is typically coherent locally and breaks globally). (a) Join…
Process Flow
Moved to references/process-flow.md to keep this file short; read it when this part is needed.
Outline file
One outline per draft target. The path is derived mechanically from output_path: same directory root, outline/, same basename, .md.
| Target | Outline |
|---|---|
output/book/text/03-methods.qmd | output/book/outline/03-methods.md |
output/article/article.qmd | output/article/outline/article.md |
It sits outside Quarto's render tree (_quarto.yml lists chapters explicitly), so it never renders. It is a living document: the architecture, plus each section's approved argument sketch, plus the per-section status. That is what makes a long chapter survive a context compaction or a break of several days — the agreements are on disk, not in the chat history. It is also what you revise against after peer review.
---
title: "<Chapter / section title>"
type: outline
target: output/book/text/03-methods.qmd
plan_task: "<Draft task id from <slug>-plan.md>"
created: YYYY-MM-DD
updated: YYYY-MM-DD
target_words: 4500
status: draft | approved
---
# Outline: <Chapter title>
## Author's formulation
<The user's own words at STOP 0, verbatim — rough is fine. Or
`agent-proposed — not yet in the author's words` if STOP 0 was overridden.
Kept as written; the sharpened version goes below.>
## Thesis
<ONE sentence, sharpened from the author's formulation: what this chapter establishes that the reader did not have
before. Not a topic ("this chapter discusses the Low Chronology") — a claim
("the Low Chronology's stratigraphic case rests on three sites whose sequences
are themselves dated by the pottery they are meant to date"). If you cannot
write it as one sentence, the chapter has no thesis yet — go back to the
synthesis pages before proposing an outline.>
## Claim chain
<2–5 sentences of prose: how the sections build from the thesis to the
conclusion, each one earning the next. Readable aloud as a chain. THIS is what
the user approves at STOP 1 — if it only makes sense as a list of topics, it is
a table of contents, not an argument.>
## Sections
### S1 — <Section title>
- **Claim:** <what this section establishes — one sentence>
- **Evidence:** `[@bibkey1]`, `[@bibkey2]`
- **Synthesis basis:** `[[synthesis-page-slug]]`
- **Function:** <what it hands to the next section>
- **Not here:** <the adjacent point that belongs elsewhere — the anti-drift line>
- **Budget:** ~<n> words
- **Status:** outlined | sketched | drafted | approved
### S2 — <Section title>
…
Section argument sketch
At STOP 2, this block goes into the outline under the section it belongs to, then gets approved. Bullets — writing prose here defeats the purpose.
#### Argument (approved YYYY-MM-DD)
- **Claim:** <the section's claim, sharpened by the discussion>
- **Steps:**
1. <step> — carried by `[@bibkey, p. XX]`
2. <step> — carried by `[@bibkey, p. XX]`
3. <step> — carried by `[@bibkey, p. XX]`
- **Concrete material:** <the specific artefact / site / passage / dataset that
illustrates the argument, named, with its source-page anchor. If you cannot
name one, this section will be assertion-stacking — find one or narrow the
claim.>
- **Counter-position:** <the strongest objection in the literature, cited> —
handled by <conceding / refuting on evidence X / bracketing as out of scope>
- **Hands to next:** <the sentence the next section can start from>
The counter-position line is not decoration. A section that cannot name what argues against it has not engaged the debate, and a reviewer will say so — cheaper to hear it here.
Reverse outline
Written after the prose, from the prose — not copied from the sketch. One line per paragraph:
#### Reverse outline — S3
| ¶ | Function | Main statement |
|---|----------|----------------|
| 1 | claim | The Negev argument rests on three sites anchored by the horizon they date |
| 2 | grounding | Tel Masos II is dated only by its pottery [@finkelstein-1999-hazor, p. 61] |
| 3 | example | … |
| 4 | counter-position | Mazar reads Arad independently [@mazar-2011-iron] |
| 5 | rebuttal | … |
| 6 | handoff | The circularity is methodological, not evidential → S4 |
Functions (use these words, so a gap is visible): claim, grounding,
example, counter-position, rebuttal / concession, definition,
method, significance, transition, handoff. A paragraph that can only be
labelled facts is knowledge telling — information lined up without being
turned towards the section's claim. Rewrite it so it states what the facts show
for the argument, or cut it.
What the user reads it for: does each line say what they mean? does the sequence follow the approved steps? is any step missing, doubled, or in a paragraph that does two jobs? The reverse outline is stored in the outline file under the section, below its approved sketch, and is updated if the prose is revised.
Citation Rules
- Inline citations:
[@finkelstein-2003]or[@finkelstein-2003, p. 152] - Multiple:
[@finkelstein-2003; @mazar-2011] - Every citation key MUST exist in
output/bibtex/references.bib - Direct quotes: inline with
>or an em-dash, always with page number - No uncited claims in argumentative sections (exception: common knowledge, clearly marked)
- Positions carry their date. Before sketching a section, run
python scripts/wiki-to-graph.py timeline <slug>for each concept or synthesis page the section argues about (MCP:graph_timeline), and read theResearch chronologysection oflint-wiki.py. The timeline lists the sources in publication order with what each says about the page and about the earlier sources; put the relevant lines into the sketch, so the order of the argument is agreed at STOP 2 and not improvised in prose. Then:- A researcher's view is what they hold in their latest source in the wiki. An earlier view they gave up (
revises) appears as history, in the past tense, with what changed their mind: "Finkelstein first dated … [@finkelstein-1996-…]; after the Tel Rehov dates he moved to … [@finkelstein-2003-…]". - A source with an incoming
supersedesis not cited as the state of the question. Cite it together with what superseded it, and say on whatbasis— new evidence dates the older position, a reinterpretation only disputes it. - Positions from different decades are not listed as parallel options unless the wiki shows the earlier one is still held (
contradicts, or aresponds-tothat leaves it open). Order them in time and name the finds or arguments that moved the discussion. - Later is not the same as right. Where a later source merely ignores an earlier objection, say so; do not present it as settled.
- A researcher's view is what they hold in their latest source in the wiki. An earlier view they gave up (
- Same author, same year, several works — citeproc disambiguates the rendered citation on its own (2026a, 2026b), so the manuscript needs nothing. What a human and the next agent need is a way to look up which is which, and that goes in a
shortlabelfield on the BibTeX entry, never innote:
The label is a fact about this manuscript's reference list, not about the work — the same title is 2026b in a project whose author wrote a different set of chapters.@incollection{hensel-2026-obadja, ... shortlabel = {Hensel 2026e}, note = {Open Access (CC BY-NC-ND 4.0)} }merge-bibs.pydropsshortlabelon merge for exactly that reason, whilenotekeeps merging because the rest of what lives there is shared fact. Putting the label innotesends one book's numbering into every other project. - Web citations (online databases, digital editions, research blogs without a BibTeX entry): use the
(domain — title)form as an inline link, e.g.[(idai.gazetteer.de — Tel Megiddo)](https://gazetteer.dainst.org/place/2048473). Use sparingly; group separately as "Web resources" in the references list.
Writing with depth (not bullet-reflow)
The single most common failure of wiki-driven drafting is bullet-reflow: each terse wiki claim becomes one flat sentence, so the prose is dense, assertion-stacked, and unreadable. Avoid it by developing every substantive point instead of merely restating it.
Three kinds of expansion — know which you are doing:
| Expansion | Allowed? | Rule |
|---|---|---|
| Grounded elaboration — examples, the explanation of why a claim holds, the argument around it | Yes, do this | Must come from the source (its page/quotes/examples or the PDF), and must be cited |
| Expository framing — transitions, restating an argument's logic, signposting | Yes | Rhetoric, not a new claim; uncited |
| New factual claim from memory | No | This is invention — the thing the wiki-only rule forbids |
Develop each substantive point — a useful pattern per paragraph (not a rigid template):
- Assertion — the claim, in your own prose (from the wiki).
- Grounding — the evidence or reasoning the source gives for it, cited
[@bibkey, p. XX]. - Example / illustration — a concrete case the source uses (a specific artefact, site, dataset, passage). This is what the wiki bullet omits and what the reader needs — reach into the source for it.
- Significance — why it matters for the section's argument (expository, uncited).
A single terse wiki bullet typically becomes a developed passage, not a single sentence. If a paragraph has assertions but no example and no explanation, it is not finished.
Reach-back procedure (checklist step 12) — escalate only as far as needed:
-
The source page's
### Direct quotesand### Examples & illustrationssections — the cheapest, already-extracted depth. -
If still too thin: open the original PDF at
<library>/pdf/<bibkey>.pdfat the page numbers the source page cites (the anchors point you straight to the passage — no full re-read), and draw out the example/explanation. -
If the source page gives no usable page anchor, do not read the PDF end to end — find the passage:
python scripts/bib-search.py "copper smelting" --key benyosef-2019-copperIt returns the physical PDF page — right for opening the file, wrong for citing. Take the printed page number from the page itself.
Search for the concept, not the string.
bib-searchis lexical: it matches the characters you typed and nothing else. A method is written differently by every discipline that borrowed it, so a single-string query measures your vocabulary, not the library. Name the method, list its aliases, run them as rank-fused--qvariants, and say which alias hit (thevia:line reports it) — see Searching for a concept. -
Cite whatever you use. If the source genuinely lacks the needed depth, say so plainly or narrow the claim — do not fill the gap from memory.
Full steps for each part are in references/writing-with-depth-not-bullet-reflow/; open the one you need:
- Searching for a concept, not a string —
references/writing-with-depth-not-bullet-reflow/01-searching-for-a-concept-not-a-string.md: Before a concept search, runpython scripts/bib-search.py index— it is incremental and costs about a second. A stale index answers about the library you had, not the one you have. - Precision rules —
references/writing-with-depth-not-bullet-reflow/02-precision-rules.md: Fluent prose is not yet precise prose. The typical failures of model-written academic text are not grammatical; they are these: - Style anchor (optional) —
references/writing-with-depth-not-bullet-reflow/03-style-anchor-optional.md: A project can anchor the prose to a voice instead of to the model's default register — the counterweight to every manuscript sounding the same. Ask the user for 2–3 pages of their own published prose (or, failing that, one article from…
MCP Optimisation (recommended)
If
dao-paper-search-mcp(seedocs/recommended-mcps.md) is available, verify citations through it instead of reconstructing them from memory. Otherwise, copy strictly from the source page's "Direct quotes" sections.
- Book / article citations:
dao-paper-search-mcp.search_crossref(query=<doi>)— check that the hit'sidentifiers.doiis the DOI you asked for — returnsinline_citation.markdown(a ready Author-Year link) andauthoritative_bibliography_line(the full references-list line). Paste both verbatim instead of formatting Author-Year yourself. - Web citations: fetch the page and classify it per
docs/source-classes.md. Ifaggregatororsuspect, either find the primary source or name the aggregator status transparently in the text.
Subagent dispatch (per section, not per chapter)
For sections over ~1200 words, dispatch the drafter subagent (see agents/drafter.md) at checklist step 11 — one section, after its sketch has passed STOP 2. Never fan out across sections: a chapter drafted in parallel cannot be steered, which is the failure this skill's staging exists to prevent.
The subagent receives:
- The approved argument sketch for this section — its contract; the prose must carry out this argument, not a reasonable-looking alternative
- The timeline of each contested concept the section treats (
wiki-to-graph.py timeline <slug>output, pasted) — which source came first, who answered whom, what has been superseded - List of synthesis pages (paths) to pull from
- List of source pages (paths) with allowed citation keys
- List of the corresponding source PDF paths (
<library>/pdf/<bibkey>.pdf— resolve withscripts/library.py) — so the subagent can reach back for examples/context when a page is thin (checklist step 11). Without these, the subagent can only bullet-reflow - The section's word budget from the outline (a floor for development, not a ceiling to pad toward)
- The project's
Manuscript styleblock from the rootCLAUDE.md, including theStyle profileif one exists
Its report includes a deviations-from-sketch line. Carry that into STOP 3 verbatim — do not quietly absorb it. A deviation the user never sees is exactly the drift the stages are meant to surface.
Quarto Template Hooks
The template's output/book/ uses a Quarto book structure (see templates/research-project-template/output/book/):
_quarto.ymldefines the chapter list — update when adding a new chapter filetext/<nn-slug>.qmdis the chapter-file naming convention (01-introduction.qmd,02-state-of-the-field.qmd, …)outline/<nn-slug>.mdholds the chapter's argument architecture — not listed in_quarto.yml, never renderedtemplate/_preamble.texholds LaTeX preamble for PDF outputMakefiletargets:make render,make preview,make clean
For articles, use output/article/article.qmd with single-file layout and output/article/outline/article.md.
Red Flags
Moved to references/red-flags.md to keep this file short; read it when this part is needed.
Key Principles
- The thesis is the author's — asked for before anything is proposed; you sharpen, check and challenge it
- Architecture before prose — the thesis and claim chain are settled and approved before a single paragraph exists
- The argument is agreed before it is written — per section: sketch, discuss, then draft. Redirecting twenty lines is free; redirecting a thousand words is not
- A stop is a stop — present and wait; never present and continue in the same message
- The outline is the memory — status per section on disk, so a long chapter survives compaction, breaks, and revision rounds
- The wiki is truth — every claim traceable to a synthesis or source page
- Elaborate from the source, not from memory — the wiki says what is claimed; the source holds the examples and explanations. Reach back and cite; a thin wiki page is a pointer, not a limit
- Develop, don't reflow — a wiki bullet becomes a developed passage (assertion → grounding → example → significance), not one flat sentence
- Deviations are surfaced, not absorbed — prose that leaves the sketch says so
- Every section comes with its reverse outline — function and main statement per paragraph, so intent can be checked without rereading
- Precise before fluent — fixed terms, connectives that mean what they assert, qualifiers kept
- Every citation verified — bibkey existence before commit
- Render check is part of drafting — not "later"
- One draft per run, one log entry — keep changes traceable
示例提示词
帮我按既定论证结构把这部分知识库内容撰写成带引文的章节初稿。
所属插件
安装该插件会一并获得同领域的 62 个 Skill / MCP。
同插件的其他科研 Skill
- 学术文献获取下载Auto-downloads open-access PDFs for graded sources and writes a manual worklist for paywalled items. Use after literature-review, before ingest-source.
- 单篇文献入库管理Adds one verified PDF to a shared bibliography library, computing its citation key and keywords. Use to register a single document outside the full workflow.
- 数字人文计算方法Surveys computational DH methods: text mining, NLP, corpus linguistics, archive standards, GIS, network analysis, and OCR. Use for an overview of available DH tools.
- 考古地理信息系统分析Runs archaeological GIS analysis with GeoPandas: catchment analysis, viewshed, kernel density, spatial statistics. Use for mapping archaeological site locations.
- 论证结构提取可视化Extracts and visualizes argument structure: Toulmin reconstruction, support/attack graphs, circularity detection. Use to diagram an argument’s logical structure.
- 考古遗物类型学分析Performs artifact analysis: typological classification, ceramic attribute coding, lithic reduction analysis. Use when classifying excavated artifacts statistically.
- 引文与出处核查Checks citations, quotations, page numbers, and bibliography entries for accuracy and source fit. Use before submission to verify source-supported claims.
- 研究选题构思设计Explores research intent, field state, methodology, and ethics, then writes an approved design doc. Use before starting a new research question or chapter.
- 术语译名一致性检查Audits a manuscript for terms, translations, abbreviations, and names whose forms must stay stable. Use when terminology drift could undermine an argument.
- 计算语用学分析Runs computational pragmatics: speech act classification, implicature detection, discourse coherence, politeness analysis. Use for language-use-and-meaning studies.
- 文献综述学术对话构建Turns verified sources into a structured scholarly conversation among positions, not a chronological summary. Use when synthesizing literature into a debate narrative.
- 语料库语言学分析Performs corpus analysis with NLTK/spaCy: collocation, keyword analysis, KWIC concordance, word2vec, diachronic trends. Use for quantitative text-corpus research.
- 数字档案馆资源检索Queries digital archive REST APIs: Europeana, Chronicling America, Internet Archive, HathiTrust, DPLA, OAI-PMH. Use when searching or harvesting heritage metadata.
- 研究项目漂移状态检测Checks a shared source library and registered projects for drift: un-indexed PDFs, unmerged keywords, bibkey collisions. Use to audit a multi-project research setup.
- EMI课堂语料转写分析Segments and annotates classroom transcripts, coding language choice and discourse function with anonymised output. Use to quantify classroom talk or translanguaging.
- 人工智能伦理分析Analyzes AI ethics: algorithmic fairness metrics, bias auditing, value alignment, moral philosophy reasoning. Use to evaluate an AI system’s ethical dimensions.
- 研究计划任务执行Executes a ready research plan task by task, gating acquisition before ingest and dispatching analysis. Use after writing-research-plan to run the planned work.
- 研究项目结项清单Runs the final completion checklist: rendering, bibliography, wiki, data, reproducibility, archiving. Use when a manuscript is finalized and ready to publish.
- 形式逻辑自动推理Applies formal logic tools: Z3 SMT solver, modal logic S5, SAT encoding, Lean4 proof verification. Use to formalize or mechanically check a logical argument.
- 研究问题凝练构建Turns a broad topic or intuition into a precise, contestable, researchable question. Use at the start of a project when no defensible question yet exists.
- 历史地理信息系统分析Performs historical GIS work: georeferencing old maps, digitizing boundaries, temporal territory overlays, population interpolation. Use for historical spatial analysis.
- 历史语言学比较分析Analyzes historical linguistics in Python: sound correspondences, cognate detection, phylogenetic trees, Swadesh lists. Use when reconstructing language relationships.
- 历史社会网络分析Builds historical network analyses: correspondence networks, prosopography, temporal community detection, Gephi export. Use to map relationships among historical actors.
- 历史文献OCR识别Transcribes printed and handwritten historical documents with Tesseract/Kraken OCR, confidence filtering, spell correction. Use for digitizing scanned historical text.
常见问题
学术手稿章节撰写能做什么?
Drafts a chapter or article section by section from synthesized wiki content, agreeing each section’s argument first. Use to turn stable research synthesis into prose。
如何安装学术手稿章节撰写?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「学术手稿章节撰写」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。
学术手稿章节撰写需要哪些权限?
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。
学术手稿章节撰写经过官方审核吗?
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。