Academic Writing Clean
Overview
Use this as the medical/biomedical manuscript writing router. It keeps the whole manuscript coherent, then dispatches section-specific work to narrower skills:
| Task |
Load this skill |
| Abstract, highlights, structured abstract |
manuscript-abstract-writing |
| Introduction / Background / gap / rationale |
manuscript-background-writing |
| Methods, reproducibility, statistics, ethics, placeholders |
manuscript-methods-writing |
| Results, figure-flow, evidence-chain narration, minimal revision |
manuscript-results-writing |
| Discussion, limitations, implications, citation-backed interpretation |
manuscript-discussion-writing |
Word .docx inspection/editing/verification |
docx-editing |
| Literature support and evidence quality |
evidence-first-research, academic-deep-search |
| Anti-AI prose cleanup |
humanizer plus the checks below |
Do not keep expanding this file into a monster. If the work belongs to one IMRaD section, load the relevant section skill.
When to Use
Use this skill when:
- The user asks for general manuscript writing, polishing, or revision without naming a section.
- Multiple sections must stay synchronized.
- A manuscript has split Markdown sources plus an integrated Word candidate.
- You need to decide which section-specific skill should own the task.
- The user asks for academic prose that is concrete, restrained, and not AI-like.
Do not use this as the only skill when:
- The task is clearly only Abstract / Background / Methods / Results / Discussion. Load the specific skill too.
- The task is file-level
.docx manipulation. Load docx-editing.
- The task requires new literature support. Load
evidence-first-research or academic-deep-search before drafting.
Global Manuscript Rules
- Separate evidence levels. Keep own data, literature, reasonable inference, and speculation visibly distinct.
- Do not invent. Never fabricate P values, fold changes, sample sizes, approval numbers, supplier/catalog information, figure panels, or citations.
- Preserve evidence boundaries. Do not collapse in vitro into in vivo, candidate into verified, association into causality, or asymmetric model arms into a single strong mechanism.
- Respect source of truth. If split Markdown files exist, treat them as authoritative unless workspace rules say otherwise. Word-only edits must be backfilled into Markdown and verified section by section.
- Prefer candidates over overwrites. Generate a clearly named candidate Word file unless the user explicitly asks to overwrite.
- Keep active directories clean. Archive scripts, logs, audits, backups, and temporary extracts under
00_inbox/agent_artifacts, 00_inbox, or the workspace's artifact directory.
- Honor minimal revision requests. When the user says “最小修改原则”, preserve the original paragraph architecture and repair only what the stated correction requires.
Default Workflow
- Identify manuscript type, target audience, target journal if known, and the requested section.
- Inspect local workspace rules (
AGENTS.md, split Markdown sources, figure maps, terminology tables) before writing.
- Choose the section skill from the dispatch table.
- If claims depend on external literature, run an evidence-first search before drafting.
- Draft or revise only the requested scope.
- Run an anti-AI and overclaim pass.
- If producing Word output, use
docx-editing validation rules: ZIP integrity, text extraction, relationship sanity when needed, and residual-marker search.
- Report what changed, what was assumed, and what remains unresolved.
Anti-AI Style Check
Before finalizing any paragraph, ask:
- Is every sentence doing work, or is it just sounding formal?
- Did I replace vague abstractions with data, actors, actions, or consequences?
- Did I avoid stock phrases such as “In the current era”, “It is worth noting”, “plays a crucial role”, and “future studies are needed” when they add no information?
- Did I vary sentence structure without making the prose theatrical?
- Did I keep the tone confident but not inflated?
- Did I remove invented certainty?
Target-Journal Style Extraction
When the user names a target journal and asks you to learn from similar published papers, do not merely summarize the papers. Build a reusable style guide or project-local skill that captures the journal's required components, article structure, evidence-boundary language, and section templates. Prefer 1–3 close articles from the exact journal plus the journal author guidelines. Keep source notes in a project-local file such as source-articles.md under that project's references directory, or an equivalent support file. See references/target-journal-style-skill.md for the full workflow.
Verification Checklist
1---2name: academic-writing-clean3description: Use when coordinating English-language academic writing and revision for medical or biomedical manuscripts. Acts as the router for abstract, background, methods, results, and discussion skills while enforcing evidence boundaries, source-of-truth rules, and anti-AI style checks.4license: MIT5---67# Academic Writing Clean89## Overview1011Use this as the **medical/biomedical manuscript writing router**. It keeps the whole manuscript coherent, then dispatches section-specific work to narrower skills:1213| Task | Load this skill |14|---|---|15| Abstract, highlights, structured abstract | `manuscript-abstract-writing` |16| Introduction / Background / gap / rationale | `manuscript-background-writing` |17| Methods, reproducibility, statistics, ethics, placeholders | `manuscript-methods-writing` |18| Results, figure-flow, evidence-chain narration, minimal revision | `manuscript-results-writing` |19| Discussion, limitations, implications, citation-backed interpretation | `manuscript-discussion-writing` |20| Word `.docx` inspection/editing/verification | `docx-editing` |21| Literature support and evidence quality | `evidence-first-research`, `academic-deep-search` |22| Anti-AI prose cleanup | `humanizer` plus the checks below |2324Do **not** keep expanding this file into a monster. If the work belongs to one IMRaD section, load the relevant section skill.2526## When to Use2728Use this skill when:2930- The user asks for general manuscript writing, polishing, or revision without naming a section.31- Multiple sections must stay synchronized.32- A manuscript has split Markdown sources plus an integrated Word candidate.33- You need to decide which section-specific skill should own the task.34- The user asks for academic prose that is concrete, restrained, and not AI-like.3536Do not use this as the only skill when:3738- The task is clearly only Abstract / Background / Methods / Results / Discussion. Load the specific skill too.39- The task is file-level `.docx` manipulation. Load `docx-editing`.40- The task requires new literature support. Load `evidence-first-research` or `academic-deep-search` before drafting.4142## Global Manuscript Rules43441. **Separate evidence levels.** Keep own data, literature, reasonable inference, and speculation visibly distinct.452. **Do not invent.** Never fabricate P values, fold changes, sample sizes, approval numbers, supplier/catalog information, figure panels, or citations.463. **Preserve evidence boundaries.** Do not collapse in vitro into in vivo, candidate into verified, association into causality, or asymmetric model arms into a single strong mechanism.474. **Respect source of truth.** If split Markdown files exist, treat them as authoritative unless workspace rules say otherwise. Word-only edits must be backfilled into Markdown and verified section by section.485. **Prefer candidates over overwrites.** Generate a clearly named candidate Word file unless the user explicitly asks to overwrite.496. **Keep active directories clean.** Archive scripts, logs, audits, backups, and temporary extracts under `00_inbox/agent_artifacts`, `00_inbox`, or the workspace's artifact directory.507. **Honor minimal revision requests.** When the user says “最小修改原则”, preserve the original paragraph architecture and repair only what the stated correction requires.5152## Default Workflow53541. Identify manuscript type, target audience, target journal if known, and the requested section.552. Inspect local workspace rules (`AGENTS.md`, split Markdown sources, figure maps, terminology tables) before writing.563. Choose the section skill from the dispatch table.574. If claims depend on external literature, run an evidence-first search before drafting.585. Draft or revise only the requested scope.596. Run an anti-AI and overclaim pass.607. If producing Word output, use `docx-editing` validation rules: ZIP integrity, text extraction, relationship sanity when needed, and residual-marker search.618. Report what changed, what was assumed, and what remains unresolved.6263## Anti-AI Style Check6465Before finalizing any paragraph, ask:6667- Is every sentence doing work, or is it just sounding formal?68- Did I replace vague abstractions with data, actors, actions, or consequences?69- Did I avoid stock phrases such as “In the current era”, “It is worth noting”, “plays a crucial role”, and “future studies are needed” when they add no information?70- Did I vary sentence structure without making the prose theatrical?71- Did I keep the tone confident but not inflated?72- Did I remove invented certainty?7374## Target-Journal Style Extraction7576When the user names a target journal and asks you to learn from similar published papers, do not merely summarize the papers. Build a reusable style guide or project-local skill that captures the journal's required components, article structure, evidence-boundary language, and section templates. Prefer 1–3 close articles from the exact journal plus the journal author guidelines. Keep source notes in a project-local file such as `source-articles.md` under that project's references directory, or an equivalent support file. See `references/target-journal-style-skill.md` for the full workflow.7778## Verification Checklist7980- [ ] Correct section-specific skill loaded for the actual task.81- [ ] Workspace rules and source-of-truth files inspected when relevant.82- [ ] If a target journal is named, author guidelines and 1–3 close source articles were checked before writing a style guide.83- [ ] No invented statistics, citations, methods metadata, or figure/panel labels.84- [ ] Evidence strength matches wording strength.85- [ ] Markdown/Word synchronization handled or explicitly not in scope.86- [ ] Unresolved items marked as `【待确认：...】` or reported separately rather than silently guessed.
ZackZ2025