---

name: paper-workflow description: >- Route general manuscript requests such as improve my paper, revise this draft, 优化论文, 润色论文 or 投稿前检查. Identify the article type, requested scope, the sections in scope and available evidence, then choose the necessary writing or checking steps. Also use when one section is named, such as rewrite the Abstract, 改摘要, 写引言 or 讨论怎么写, or when a draft must reach a broad reader (面向大同行, 写给非本领域读者, too technical), so that the section's contract, its reader and its dependencies on other sections apply. Also use when prose reads as AI-generated (去 AI 味, AI 味重, 像 AI 写的, sounds AI-generated). Go to a specialist directly when the user names one or a narrow job such as a citation style. Do not route ordinary emails, code documentation or non-manuscript writing through this workflow.

Paper Workflow: Dispatcher

Use this dispatcher when the author has not identified a specific writing task. A request such as 优化一下论文 or improve my manuscript names a goal, not a layer. Manuscript work happens at distinct layers, and editing the wrong layer first wastes the edit: a paragraph whose scientific role is wrong should never be polished, and a sentence whose claim is unstable should never be re-punctuated.

Classify the request before editing. Check the relevant layers in scientific-to-stylistic order, then load and apply only the specialists needed to resolve observed problems. A single specialist may be sufficient after diagnosis; a general request does not itself require repeated rewriting.

Default assumption: unless a conference venue is named, the manuscript follows the journal-oriented Nature-style path.

Where the rules come from

Every rule in these skills serves a reader of the paper at its target venue: the editor who decides whether it goes to review, the reviewer who checks the evidence, or the reader who uses the result. The rules are written for the default case, a research article in a Nature Portfolio journal. Before applying a rule, know whose need it serves. When the venue, the article type or the reader differs, derive the rule again from that need and say so in the reply. Text that meets a rule's wording while defeating its purpose fails the rule. When a rule seems to work against its own purpose in a case, report it rather than follow it silently.

Step 1: Establish the task and classify the input

Identify the active source, article type, requested scope, the positions in scope (Abstract, Introduction, Results, Methods and so on; see Positions below) and output before choosing a route. Use supplied context; ask only when an ambiguity would change the work. A request to inspect, review, diagnose or suggest produces findings, not file edits. A request to revise authorizes edits within its stated scope. Explicitly frozen decisions and numerical results remain fixed.

Name the target venue, then ask two questions from its side: what do its editors and reviewers check, and who reads it? The answers decide what the paper must say and what it can leave out. A sentence that one venue's reviewers look for can read as defensive at another: an ML conference needs a plain statement that test data played no part in model selection, which a journal edit might cut. nature-portfolio-playbook covers Nature Portfolio journals and conference-paper-writing covers ML conferences. When no venue is named, take the default journal path and say so.

Carry forward named protected passages and the requested edit depth. A light-touch request does not reopen approved Abstract, Introduction or Discussion prose. If a protected passage contains a material scientific error, report it with a concrete remedy; preserving its text is not a clean bill of health. Reuse supplied constraints rather than demanding a new state file.

Treat the byline, correspondence and contribution roles as author decisions. Checking initials or aligning contribution formatting does not authorize adding/removing authors or changing their roles. Resolve initials against full names and current source; flag an ambiguous mapping rather than infer an identity or propagate a suspected correction into other metadata.

Work with the material provided. Do not demand project-state files for a paragraph edit. For whole-paper work, reuse existing results and decision records; create lightweight notes only when they prevent cross-session drift. If the input is a PDF, determine whether editable source exists; do not promise Word track changes, PDF layout editing or LaTeX compilation without a working toolchain. For missing evidence, identify the affected claim and continue with independent edits. Never fabricate the missing result or citation.

The table gives candidate steps in order, not mandatory rewrite passes.

Input	Class	Chain
One sentence or one paragraph	passage	write-scientific-manuscript, then anti-defensive-writing if the passage is hedged, over-caveated or written in an audit, self-critical, developer or commitment voice, then scientific-prose-style
One section to draft or rewrite in prose, such as the Abstract or the Introduction	section	Read that position's entry in references/section-contracts.md and check its upstream dependency; scientific-writing for needed drafting; passage, posture and sentence specialists only for observed issues
Written for specialists but must reach a broad reader: 面向大同行, 写给非本领域读者, "too technical"	audience	Read the Readers section of references/section-contracts.md; manuscript-optimizer for framing in the broad positions (the reader's problem by the second sentence, what success is measured against, findings in the reader's terms), then write-scientific-manuscript for terms and passage logic, then scientific-prose-style. Technical positions keep their precision
Reads as AI-generated: 去 AI 味, AI 味重, 像 AI 写的, "sounds AI-generated"	ai-tone	Run scientific-prose-style's scripts/ai_tells.py on the text in scope and read its references/ai-tells.md. Then write-scientific-manuscript for coined terms and weak subjects, anti-defensive-writing for recaps, forward pointers and rhetorical questions, and scientific-prose-style for words, frames and typography. Apply the information test: delete what carries nothing, never swap synonyms, never cut a number
A Results section that is scientifically settled but reads figure-by-figure	results-flow	results-section-revision, anti-defensive-writing, then scientific-prose-style
A whole draft, or no unit named	manuscript	manuscript-optimizer, with references/section-contracts.md as the per-section standard; scientific-writing for needed drafting; passage, posture and sentence specialists only for observed issues
Reads like an audit report, a rebuttal or a self-critique; hedged, over-caveated or apologetic; a paragraph opens with a limitation; the SI or data statements confess what could not be found; the text names folders, files or build steps, or promises a future release	posture	anti-defensive-writing, then scientific-prose-style. If the claim hierarchy is not yet settled, run manuscript-optimizer first: an unnecessary disclaimer and a real scope condition look identical while the claim is still moving
A Review, survey, or Perspective	review-article	review-article-architecture first, then the Review path below
A draft carried across many sessions	long-draft	Reconcile current source and decisions with draft-marker-discipline if markers/state are relevant; research articles use manuscript-optimizer, Reviews/surveys/Perspectives use review-article-architecture. Length alone never selects the Review route
Near submission or resubmission	preflight	submission-audit, citation-verifier, claim-source-verification, stats-reporting-audit, data-availability
Reviewer comments exist	response	paper-reviewer to inventory every ask, rebuttal-response to draft and calibrate, then paper-reviewer again to grade the draft. paper-reviewer and rebuttal-response are both in the recommended set
A manuscript to referee, or a request to predict what reviewers will attack	referee	paper-reviewer (recommended set)
Figures are the bottleneck	figure	figure-planner, then nature-figure to render, then figure-style to check. The last two are the figure stack, installed with --figure
Equation numbering, float placement, SI table pagination or other LaTeX typesetting	layout	scientific-writing with references/latex-layout.md; use write-scientific-manuscript only if paragraph boundaries also need semantic repair. Placement alone does not select the figure-production chain
No draft yet, project new or messy	bootstrap	paper-bootstrap then nature-portfolio-playbook

An explicit mixed request can use the relevant steps from more than one class. Ask one question only when competing interpretations would change the authorized work; do not ask merely because numbering, paragraphs and floats need different specialists.

Step 2: Announce the scope and execute the necessary steps

State the chosen scope and sequence in one short line. For example:

I will check the claim/evidence structure, revise the affected Results paragraphs, and check the changes against the figures and original numbers.

Check the layers in order, but distinguish checking a layer from rewriting it. If a layer is already sound, preserve it. Do not reopen settled structure during a sentence edit. Apply a posture edit only after the relevant claim is stable. When a requested comprehensive preflight cannot check a category, report that category as unchecked and explain what input is missing.

Finish each edit with a focused check of the changed text and its dependent figure references, comparisons, terminology and citations. Do not repeatedly rewrite stable prose to satisfy a chain.

Output contract

• Revision: deliver the revised text/file first, then at most 3–5 material changes and any unresolved author decisions. A short passage normally needs only the replacement and a brief reason.
• Diagnosis: give prioritized findings with locations, reasons and concrete remedies; stop before edits.
• Preflight: distinguish blocking issues, useful improvements and unchecked categories. Do not turn missing tools or unread sources into a clean bill of health.
• Rebuttal: preserve every reviewer ask and original numbering; link replies to actual manuscript changes and completed evidence. A proposed experiment is not a completed experiment.
• Never claim an export, render, citation lookup, experiment or file modification was performed unless it was performed. Keep detailed process notes outside the manuscript.

Step 3: Stop rules

Stop or narrow the affected work and explain why when:

• an upstream layer finds a problem that invalidates downstream work, such as a claim the evidence does not support. Fix or surface it before polishing;
• the requested scope is complete. 帮我把这句话改短 ends after the passage edit and its local check;
• a later layer would undo a decision the author explicitly approved.

Run a full preflight once the draft is stable; if the author requests it earlier, identify instability and check only categories that can be assessed. Auditing unstable text produces findings that evaporate on the next revision.

Never run scientific-prose-style on a drifted Review before the drift audit. Polishing a drifted draft makes the drift harder to see, not easier.

Never run anti-defensive-writing before the claim hierarchy is stable. Until the claim is settled, an unnecessary disclaimer and a real scope condition look identical, and the pass strips the wrong one.

The layers

Use these layers to diagnose problems and preserve dependency order.

1. Structure: is there a defensible claim hierarchy and evidence chain, framed for the intended reader? manuscript-optimizer for research articles, review-article-architecture for Reviews.
2. Prose: is the section drafted in full paragraphs that carry the argument? scientific-writing, with results-section-revision for late-stage Results architecture.
3. Passage logic: is this paragraph followable? Buried topic sentences, missing bridges, ambiguous referents, noun chains, incomplete comparisons, coined terminology. write-scientific-manuscript.
4. Rhetorical posture: does the text advance its claim, or negotiate with an imagined critic? Audit voice (project process in the paper), defensive voice (arguing with an imagined reviewer), self-critical voice (the paper grading itself), developer voice (the repository in the paper), commitment voice (promises of a future release), caveats in high-impact positions, paragraphs that open with a limitation, reflexive not X but Y. anti-defensive-writing, which also covers the SI, legends and data statements. It runs after the claim hierarchy is settled, because before that an unnecessary disclaimer and a real scope condition are indistinguishable: the claim they qualify is still moving.
5. Sentence: em-dash and semicolon budget, hedging, sentence rhythm, paragraph openers. scientific-prose-style, last.

Integrity checks run alongside, not in sequence: citation-verifier, claim-source-verification, stats-reporting-audit, draft-marker-discipline.

Layer 4 has a boundary the other layers do not: a load-bearing statement stays. Load-bearing is decided by what a sentence states, not by which skill added it: the unit of replication and n, no P values, blinding and exclusions, an access route, the count behind a reporting decision, an estimator's assumptions, a null result, any scope condition the Methods needs to stay reproducible, and any statement the target venue's reviewers look for. anti-defensive-writing may move such a statement out of a high-impact position, state it once instead of at every mention, or rewrite it as positive scope, but it must not delete it. A pass that silently removes one is a reporting failure, not a style improvement. Audit-voice wording added by an integrity check ("as checked", "could not be identified") is not protected.

Positions

The layers decide what to fix first. The position decides what a fixed section must do. An Abstract, an Introduction and a Methods section are held to different standards, and some layers behave differently in them: a Methods section is almost entirely load-bearing, and an Abstract keeps effect sizes but not test statistics. section-contracts.md states each position's job, dependencies, acceptance criteria and owning specialists once. Specialists apply it and do not restate it. Positions also differ in register: the front matter and the opening and close of each section are written for the broad peer (大同行), the body of Results and Methods for specialists. The contracts' Readers section defines both and how to test them. Read the entry for every position in scope before diagnosing it, and check the edit against the same entry afterwards.

Positions depend on one another:

figures and evidence → Results → Discussion → Introduction → Abstract → Title

Methods records what was actually done and changes only with the procedure. Three rules follow:

• A request for a downstream position stays narrow. Edit it against the current upstream text, keep claims that rest on unsettled results out of it or mark them as pending, and name the dependency in the reply.
• After a claim changes upstream, recheck the positions that restate it: the Discussion opening, the final Introduction paragraph, the Abstract, the Title and the affected legends. This is a focused check, not a rewrite.
• No position claims more than its upstream supports.

Format is not part of a contract. Abstract length and structure, citations in the abstract, title length and Methods placement come from the target journal's current guide, through nature-portfolio-playbook for Nature Portfolio journals. A generic template, such as a structured clinical abstract, is never the default.

Default journal path

Use this as a coverage map over the life of a paper; run only the steps applicable to the current request.

1. paper-bootstrap
2. nature-portfolio-playbook when venue fit or article type is uncertain
3. Refresh notes/project_truth.md, notes/result_summary.md, notes/paper_handoff.md after any experimental, statistical, or figure update
4. manuscript-optimizer or scientific-writing per the class table, then write-scientific-manuscript for passage-level clarity
5. figure-planner, then nature-figure to produce and figure-style to check (figure stack, --figure)
6. results-section-revision when Results is stable but reads jumpy
7. stats-reporting-audit
8. citation-verifier, then claim-source-verification
9. data-availability
10. anti-defensive-writing, after the integrity checks above have added the statements they require
11. scientific-prose-style
12. submission-audit; anything it adds is held to rule zero of anti-defensive-writing, and that skill's detection pass is rerun on the changed passages
13. after external review: paper-reviewer to inventory the reports, rebuttal-response to draft and calibrate, then paper-reviewer again to grade the draft

Review, survey, and Perspective path

A Review is not a short research article. Its failure mode is becoming a different Review, not overclaiming past its data.

1. review-article-architecture to establish the governing plan document first
2. nature-portfolio-playbook for venue and article-type fit
3. draft-marker-discipline to set up the marker system before drafting starts
4. scientific-writing, sourcing in step with the prose
5. citation-verifier, then claim-source-verification
6. figure-planner, then nature-figure and figure-style (figure stack, --figure)
7. review-article-architecture drift audit, before any compression pass
8. draft-marker-discipline to measure length and triage what remains open
9. anti-defensive-writing
10. scientific-prose-style, last
11. submission-audit

Run step 7 before step 10, never after.

Choosing between adjacent skills

• scientific-writing when a section mostly needs drafting or rewriting in prose, or when a citation style or reporting guideline is named. manuscript-optimizer when the story, evidence chain, figure logic, or terminology may be unstable. write-scientific-manuscript when the science is settled but the passage is hard to follow.
• results-section-revision when the remaining problem is local Results architecture rather than claim selection.
• anti-defensive-writing when the text is accurate but reads like an audit report, a rebuttal or a self-critique: it reports project process instead of findings, keeps saying what it does not claim, opens paragraphs with caveats, confesses what could not be found, explains itself to a critic who is not in the room, documents the repository, or promises what does not exist yet. It covers the SI, legends and data statements as well as the main text. scientific-prose-style when the remaining problem is punctuation, rhythm, em-dash budget, or hedge calibration inside a single sentence. They stack, in that order, because removing defensive scaffolding rewrites the paragraph openers and sentence boundaries the punctuation pass then settles. Neither one may overrule an integrity audit; see the layer-4 boundary above. A Methods section is mostly load-bearing reproducibility detail, and the edit is usually to convert a negation into positive scope and to state it once instead of at every mention.
• citation-verifier when the bibliography as an artifact is the problem: duplicate keys, missing fields, DOI syntax, cited-but-undefined. claim-source-verification when the question is whether a source supports the sentence citing it. They stack, in that order. A clean bibliography audit says nothing about claim support: in one measured run, 55 of 139 proposed sources were rejected, none of them fabricated.
• reference-audit-guide when references must be checked against live scholarly APIs rather than inspected locally. Optional set, installed with --set all. It ships runnable verification scripts and is the only stage that catches a fabricated citation carrying a well-formed DOI.
• review-article-architecture for a Review, survey, or Perspective, or whenever a piece written across many sessions may no longer match its brief. manuscript-optimizer for research articles.
• draft-marker-discipline before a batch pass over open markers, before quoting a manuscript's length, before removing superseded material, and before scripting the same edit across many files.
• figure-planner to decide what each figure argues, then nature-figure to render it, then figure-style to check correctness and legibility before export. The last two are the figure stack; if they are not installed, figure-planner still produces the panel plan and the author renders it themselves.
• conference-paper-writing only when a conference venue is explicitly named. Optional set, installed with --set all.
• paper-reviewer when the question is what the reviewer asked or whether the reviewer would accept the answer. rebuttal-response when the question is what the authors may claim and how the letter, manuscript, and Supplementary Information stay consistent. They stack, in that order, and paper-reviewer runs a second time at the end to grade the draft. rebuttal-response is in the recommended set; paper-reviewer is in the recommended set. A manuscript change made in reply to a reviewer follows rule zero of anti-defensive-writing: a result, a precise definition or a narrowed claim, not an added caveat.

Working principle

Do not polish sections written from stale memory. After experimental, statistical, or figure updates, refresh notes/project_truth.md, notes/result_summary.md, and notes/paper_handoff.md before attempting heavy revision.

Common mistakes

• editing a general manuscript request without first identifying the actual problem layer
• treating the coverage map as a requirement to rewrite every section on every request
• polishing sentences before the claim hierarchy is stable
• writing the Abstract or Title from a Results claim that is still moving, or failing to recheck them after it moves
• importing an abstract format, such as a structured clinical abstract, that the target journal does not ask for
• holding every section to one generic standard instead of its own contract
• running a defensive-writing pass before the claim hierarchy is stable, which strips real scope conditions along with the disclaimers
• letting a posture edit delete scientifically necessary limitations, numerical results or reproducibility facts
• running a submission audit on a draft still being restructured
• polishing a Review before its drift audit
• using conference-style writing skills by default for journal manuscripts
• deciding what to keep without asking what the target venue's reviewers check, for example cutting an ML paper's no-leakage statement as defensive wording
• answering an evaluation-design problem with a label such as "prespecified" instead of reporting the result the reviewer would ask for
• treating AI tone as a word list: swapping synonyms or cutting common words such as "robust" instead of deleting what carries no information
• rewriting the manuscript from experiment memory instead of a current result_summary.md
• editing figure legends late without rechecking the Results text
• leaving repository choice, accession IDs, or source-data coverage until the portal is open
• treating citation formatting as the same thing as citation verification
• treating a clean bibliography as evidence that the sources support the claims
• writing a response letter before deciding the underlying manuscript edits
• answering a reviewer report without first listing every ask and sub-ask in the reviewer's own words
• renumbering, reordering, or grouping reviewer comments by theme, which makes coverage impossible to check at a glance
• rewriting whole paragraphs when the author flagged two sentences