Academic Editing

Invocation Model (Single User Entry Point) / 调用方式（单入口）

Users should invoke only academic-editing. The files under references/ are internal docs, not standalone skills. The skill automatically chooses the needed internal mode; see the Mode Map below.

Start by identifying the article/study type, then load the matching reference(s).

Mandatory Intake Before Writing / 写作前必做信息收集

Ask only what is needed to proceed.

• Always ask: the primary scientific question/title, article/study type, and target journal (or top 2 candidates)
• For fact-heavy or full-manuscript work, also ask where the code, outputs/tables, and figures live; if absent, use /code, /output, and /figure
• If the user already provided any item, do not ask it again

Study-Type and Article-Type Routing (Internal references) / 研究类型与文稿类型路由（内部参考）

• Clinical cohort / observational studies: references/cohort-study-edit.md
• Bioinformatics studies (GWAS, single-cell, MR): references/bioinfo-edit.md
• Narrative reviews, perspectives, review-style articles, title/abstract rewriting, and journal-tone polishing: references/journal-writing.md
• Clinical manuscripts, biomedical prose, and fact-heavy revisions after analysis: references/clinical-accuracy.md
• Full papers, major rewrites, long reviews, and heavily revised drafts: references/manuscript-coherence.md
• If mixed design, combine the relevant references and enforce one unified manuscript voice.

Evidence-First Rule (Non-negotiable) / 证据优先规则（不可协商）

Before rewriting claims that include numbers, model results, dates, sample sizes, effect estimates, or causal language:

1. Check the corresponding code/output/figure evidence first.
2. Verify consistency across text, tables, and figures.
3. If evidence is missing or inconsistent, flag it explicitly before final polish.

Reporting-Guideline Rule / 报告规范规则

Always map the study design to a reporting guideline before polishing:

• Trials and protocols: CONSORT / SPIRIT
• Observational or cohort studies: STROBE
• Reviews and meta-analyses: PRISMA
• Diagnostics and prediction: STARD / TRIPOD, with AI extensions when relevant
• Case reports, QI, and guidelines: CARE / SQUIRE / RIGHT or AGREE as required
• MR, GWAS, and other omics: STROBE-MR / STREGA / the relevant domain guideline plus journal and repository requirements

If guideline requirements conflict with local constraints, explain the conflict and apply the safer compatible choice.

Read: references/guideline-sources.md.

Editing Order / 写作顺序

1. Evidence lock first: verify numbers, model results, dates, sample sizes, effect estimates, and causal claims against code, outputs, or figures.
2. Map the study type to a reporting guideline, then load the matching reference(s).
3. Apply the relevant writing pass: journal-writing for paper-ready prose and review-style text; clinical-accuracy for clinical or fact-heavy revisions; manuscript-coherence for full papers, long reviews, or major rewrites.
4. If wording changes facts or cross-section logic, rerun Loop.

Mode Map / 模式地图

These are routing rules, not user commands.

Common Mistakes / 常见失误

• "Polish first, verify later" -> evidence lock first; if a factual sentence cannot be traced, flag it.
• "The numbers are probably close enough" -> re-check the source data, outputs, and tables until the exact value is confirmed.
• "This was only a wording change" -> if claims, labels, or cross-section logic may shift, rerun Loop.
• "A long paper can be fixed section by section" -> build the manuscript map first and confirm one thesis, one strength level, and one journal fit.
• "The introduction or discussion can be improved in isolation" -> confirm the section still supports the scientific problem, gap, novelty, and take-home message.

If the same evidence mismatch or coherence gap appears twice, stop polishing and return a short issue block with the exact mismatch and missing source evidence.

Smoke Test / 烟测

• "Please polish this Results paragraph." -> direct evidence + style pass
• "Please revise this full review for cohesion." -> PI, then Loop if wording shifts facts
• "Please polish this clinical Discussion and verify numbers." -> evidence lock + clinical-accuracy + Loop after revision

Internal Modes / 内部模式

Think in two internal modes rather than requiring user-facing commands:

• PI (Principal Investigator / 总负责人): for long manuscripts, reviews, and multi-section rewrites. PI owns the whole manuscript, delegates bounded subtasks to subagents when useful, and keeps the final narrative, evidence, and quality bar consistent.
• Loop (Revision Loop / 修订循环): for post-polish fact checks, claim drift checks, and repeated consistency sweeps.

These modes are automatic behavior, not extra commands the user needs to type.

PI execution pattern:

• Define the manuscript objective, target journal fit, and the section-level work list.
• Split bounded subtasks only when they are independent and do not require sequential context.
• Delegate fact checks, table/figure tracebacks, or section-specific rewrites to subagents only when that reduces risk and preserves evidence-first control.
• Recombine outputs into one coherent manuscript voice before any final polish.
• Hand the result to Loop if the wording change may have shifted numbers, claims, or cross-section consistency.

Loop execution pattern:

• Re-scan any section whose wording changed materially.
• Re-check numbers, labels, causal language, and section-to-section consistency.
• Stop only when the revised text still matches the underlying evidence and the full manuscript narrative.

Style and Language Requirements / 语言与风格要求

• Rephrase for clarity, coherence, and conciseness.
• Ensure paragraph-to-paragraph logical flow.
• Use a professional SCI tone.
• Remove unnecessary jargon.
• Keep terminology consistent across the full manuscript.

Output Contract / 输出约定

• Default output: polished text only.
• Highlight only words/sentences actually revised using bold (**...**).
• Do not add unnecessary commentary.
• Exception: if a scientific/logical inconsistency is detected, output a short issue block first:

[Issue] - Problem: ... - Evidence mismatch: ... - Required fix before polishing: ...

Then provide the polished text.