/paper-workflow

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 `s