Paper Workflow

Paper Workflow: Dispatcher

This is the entry point, not a menu. 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.

Do not answer a general manuscript request by loading a single specialist skill. Classify the request, then run the whole chain for that class.

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

Step 1: Classify by what was actually handed over

When two classes with different chains both fit, ask one question. That is the one case worth a clarifying question; guessing wastes more time than asking.

Step 2: Announce the chain, then run it in order

State the chain in one line before starting, so the author can redirect early:

Running the manuscript chain: structure (manuscript-optimizer), prose (scientific-writing), passage logic (write-scientific-manuscript), posture (anti-defensive-writing), then sentence pass (scientific-prose-style).

Load each skill in sequence and apply it. Do not skip a link because the previous one already improved the text. Each layer catches a different defect class, and a later layer cannot see the defect an earlier one owns.

Step 3: Stop rules

Stop the chain and report 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 request was narrow. 帮我把这句话改短 is a passage job, not a licence to restructure the paper;
• a later layer would undo a decision the author explicitly approved.

Never run preflight on a draft still in manuscript class. 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

Every skill sits at one layer. This is why the chains are ordered.

1. Structure: is there a defensible claim hierarchy and evidence chain? 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? Unnecessary disclaimers, caveats in high-impact positions, paragraphs that open with a limitation, repeated statements of what the text does not claim, reflexive not X but Y, self-undermining contribution statements. anti-defensive-writing. 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 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 limitation placed by an integrity check is load-bearing and stays. anti-defensive-writing may move it 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. The same holds for a scope condition the Methods needs to stay reproducible. A defensive-writing pass that silently removes a mandated caveat is a reporting failure, not a style improvement.

Default journal path

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 placed every caveat they require
11. scientific-prose-style
12. submission-audit
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 timid: it keeps saying what it does not claim, opens paragraphs with caveats, hedges what the evidence actually supports, or explains itself to a critic who is not in the room. 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. In a Methods section, most caveats are load-bearing, 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. Optional set, installed with --set all.

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

• answering a general manuscript request with whichever single skill matched the wording best
• polishing sentences before the claim hierarchy is stable
• running a defensive-writing pass before the claim hierarchy is stable, which strips real scope conditions along with the disclaimers
• letting a defensive-writing pass delete a limitation that a statistics, citation, or data audit put there, instead of moving it out of a high-impact position and stating it once
• 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
• 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