AgentSkillsCN

scientific-writing

每当用户要求撰写科学论文、文献综述、研究论文、病例报告、系统性综述,或任何学术/研究类文档时,都可使用此技能。触发短语包括:“撰写论文”、“撰写文章”、“文献综述”、“综述文章”、“病例报告”、“研究论文”、“科学论文”、“学术论文”,或要求基于科学文献撰写各类文档时。此技能会强制执行结构化的科研工作流程,包括来源筛选、循证写作,以及清晰披露哪些文献被完整阅读,哪些仅查阅摘要。此外,当用户要求将多篇科学论文进行总结、综合或对比,最终形成书面交付成果时,也应使用此技能。

SKILL.md
--- frontmatter
name: scientific-writing
description: >
  Use this skill whenever the user asks to write a scientific article, literature review,
  research paper, case report, systematic review, or any academic/scholarly document.
  Triggers include: "write a paper", "write an article", "literature review", "review article",
  "case report", "research paper", "scientific article", "academic paper", or requests to
  produce documents citing scientific literature. This skill enforces a structured research
  workflow with source triage, evidence-based writing, and transparent disclosure of which
  sources were read in full vs. abstract only. Also use when the user asks to summarize,
  synthesize, or compare multiple scientific papers into a written deliverable.

Scientific Article Writing Skill

Core Principle

Abstracts are valid sources. Be transparent about what you used.

Abstracts contain legitimate, citable information: quantitative results, study design, sample sizes, effect sizes, p-values, and main conclusions. If the abstract provides the data you need, there is no obligation to retrieve the full text — doing so consumes context without adding value.

However, transparency is non-negotiable. The user must always know which sources were read in full and which were used via abstract only. Before presenting the final document, include a clear disclosure listing abstract-only sources (see Source Transparency below).

Full-text reading remains valuable for interpretive claims, methodological details, subgroup analyses, and limitations not covered in the abstract. Use your judgment: if the abstract is sufficient, use it; if you need more depth, attempt full-text retrieval.

PDF Upload Policy

PDF upload to Zotero happens only when the user explicitly requests it:

  • If the user explicitly requests PDF upload → it is mandatory and cannot be skipped. Follow the verify-upload-validate procedure in the zotero-mcp-integrations skill. For paywalled sources, present a table to the user asking them to upload the PDFs manually if they have access.
  • If the user does NOT request PDF upload → do not upload PDFs.

Source Transparency (MANDATORY)

Before presenting the final document, you MUST disclose which sources were used via abstract only. A simple list is sufficient:

Sources used via abstract only: Smith et al. 2021, Jones et al. 2023. All other sources were read in full.

This is not a warning — it is an informational disclosure that lets the user judge the evidence quality and decide whether to provide additional full texts.


Workflow Overview

The workflow has 5 mandatory phases executed in strict order. No phase can be skipped. Each phase has explicit entry/exit criteria.

code
Phase 1: SCOPE → Phase 2: SEARCH → Phase 3: READ → Phase 4: WRITE → Phase 5: DELIVER

See references/workflow-detail.md for the complete phase specifications.


Phase 1: SCOPE — Define the Article

Goal: Understand exactly what needs to be written before searching anything.

Actions:

  1. Identify article type (review, research, case report, systematic review, meta-analysis, editorial, letter, etc.)
  2. Determine target audience and field
  3. Determine approximate length and section structure
  4. Identify citation style (Vancouver, APA, IEEE, etc.)
  5. Clarify key questions the article must answer
  6. Agree on scope boundaries — what is IN and what is OUT

Exit criteria: Clear written specification of what will be produced.

Anti-pattern: Starting to search before knowing what you're looking for. This leads to unfocused literature retrieval and wasted effort.


Phase 2: SEARCH — Find Sources

Goal: Build a comprehensive, relevant bibliography.

Actions:

  1. Construct search queries based on the scope (use PubMed, web search, or whatever tools are available)
  2. Cast a wide initial net — search for more sources than you'll ultimately cite
  3. Triage results: categorize by relevance (essential / supporting / background)
  4. Record all sources with their identifiers (DOI, PMID, URL)

Search strategy by article type:

Article TypeMinimum SourcesSearch Depth
Narrative review15-30Broad, thematic
Systematic reviewExhaustive per protocolProtocol-driven, reproducible
Research article10-25Focused on methods + context
Case report5-15Similar cases + guidelines
Short communication5-10Targeted, recent

Exit criteria: A ranked list of sources with abstracts (and full text where easily available).

Anti-pattern: Stopping at 10 PubMed results.


Phase 3: READ — Understand the Sources

Goal: Develop sufficient understanding of the source material before writing.

When abstracts are enough

For many sources — especially supporting or background references — the abstract provides all the data you need: study design, sample size, key results, main conclusion. In these cases, there is no need to retrieve the full text.

When full text adds value

For essential sources where you need interpretive depth, methodological detail, subgroup analyses, or limitations, attempt full-text retrieval:

  1. web_fetch on DOI URL — often resolves to full HTML text on publisher sites
  2. Open access repositories — PMC, Europe PMC, BioRxiv, MedRxiv
  3. Publisher open access — JMIR, PLOS, BMC, Frontiers, MDPI are fully OA
  4. Preprint versions — BioRxiv, MedRxiv, arXiv, SSRN

If full text is behind a paywall, the abstract is sufficient. Do not waste effort trying to circumvent access restrictions.

What to extract from each source

Whether from abstract or full text, for each source you should know:

  • What was the study question?
  • How was it designed? (sample size, population, key methods)
  • What were the main numerical findings?
  • How does this fit into the broader evidence base?

For full-text sources, also extract:

  • Methodological details not in the abstract
  • Stated limitations
  • Subgroup analyses or secondary endpoints

Track your source depth

Keep a mental note of which sources you read in full and which you used via abstract only. You will need this for the Source Transparency disclosure in Phase 5.

Exit criteria: Sufficient understanding of each source to make accurate claims.

Anti-pattern: Paraphrasing abstract conclusions without extracting specific data. Even when using abstracts, cite concrete numbers, not vague summaries.


Phase 4: WRITE — Compose the Article

Goal: Produce a well-structured scientific article grounded in the sources you've read.

Structure by Article Type

See references/article-structures.md for detailed templates.

General principles that apply to all types:

  1. Introduction: Establish context → Identify the gap/need → State the objective
  2. Body: Organize by theme, chronology, or methodology — NOT by source. Never write "Smith et al. found X. Jones et al. found Y. Lee et al. found Z." Instead: "Multiple studies have demonstrated X [refs], although the effect varies by population [ref1, ref2]."
  3. Discussion/Synthesis: Integrate findings, identify patterns, acknowledge conflicts
  4. Conclusions: Answer the question posed in the introduction

Writing Rules

  • Every factual claim needs a citation. No exceptions.
  • Use specific data, not vague summaries. Say "92.5% maintained viral suppression at 48 weeks" not "most patients responded well." Abstracts often contain these specific numbers — use them.
  • Acknowledge contradictions in the literature. Don't cherry-pick.
  • Distinguish between what the data shows and what it suggests. Use language appropriately: "demonstrated" vs "suggested" vs "may indicate."
  • State limitations explicitly, both of individual studies and of the review itself.
  • Track citation numbers carefully if using numbered styles (Vancouver, IEEE). Each unique source gets one number, assigned in order of first appearance.

Citation Placement

  • Citations go AFTER the claim, BEFORE the period: "...maintained suppression [4]."
  • Multiple citations: [4,5] or [4-6] for ranges
  • Abstract-only sources can be cited for specific factual claims directly stated in the abstract (quantitative results, study design, sample size, main conclusion). Do NOT use abstract-only sources for interpretive or methodological claims that require full-text context.

Exit criteria: Complete draft with all citations placed.

Anti-pattern: Writing the full text and then "sprinkling" citations afterward. Citations should be integral to the writing process, not a post-hoc decoration.


Phase 5: DELIVER — Produce the Final Document

Goal: Generate a polished, properly formatted deliverable.

Actions:

  1. Generate the document in the requested format (.docx, .pdf, .md, etc.)
  2. If Zotero integration is available, inject citation field codes
  3. Include the Source Transparency disclosure (see Core Principle) listing which sources were used via abstract only
  4. If the user requested PDF upload, present a table of paywalled sources they may want to upload manually (see zotero-mcp-integrations skill)
  5. Validate the document
  6. Present to user with clear instructions for any post-processing (e.g., Zotero Refresh)

Exit criteria: Downloadable document with working citations.


Tool-Agnostic Principles

This skill works regardless of which tools are available. The workflow is the same whether you have:

  • PubMed MCP tools + Zotero MCP tools + web_search
  • Only web_search + web_fetch
  • Only web_search
  • No search tools at all (user provides sources)

The key insight is: the workflow is about the PROCESS, not the tools.

If you have...Use it for...
pubmed_search_articlesPhase 2: Structured literature search
pubmed_fetch_contentsPhase 2-3: Get abstracts and metadata
web_searchPhase 2: Find sources; Phase 3: Locate full text when needed
web_fetchPhase 3: Read full-text HTML when abstract is not sufficient
import_pdf_to_zoteroPhase 5: Archive PDFs when user requests it (see zotero-mcp-integrations skill)
find_and_attach_pdfsPhase 3/5: Batch OA PDF lookup + auto-attach via Unpaywall
get_item_fulltextPhase 3: Read indexed full text + validate post-upload content
add_linked_url_attachmentPhase 5: Fallback when PDF import fails
inject_citations / skillPhase 5: Create live Zotero citations in .docx
add_items_by_doiPhase 2: Build bibliography in Zotero
add_itemsPhase 2: Add books, theses, reports, etc. that do not have a DOI
None of the aboveAsk the user to provide source material directly

Zotero integration: If the zotero-mcp-integrations skill is available and the user requests PDF upload, follow its Step 2b during Phase 5. For citation injection, follow its Steps 4-6 during Phase 5.


Quality Checklist

Before delivering the final document, verify:

  • Every factual claim has a citation
  • Specific numbers (percentages, sample sizes, CIs) are cited — abstracts are a valid source for these
  • Contradictions in the literature are acknowledged
  • Limitations are stated
  • The article answers the question defined in Phase 1
  • Citations are numbered/formatted correctly for the chosen style
  • Abstract-only sources are used appropriately (factual claims only, not interpretive)
  • Source Transparency disclosure was presented listing all abstract-only sources
  • If user requested PDF upload: freely available PDFs were verified and uploaded; paywalled sources listed in a table for the user
  • No PDF was uploaded without prior content verification (title, authors, actual content)

Common Failure Modes

FailureCausePrevention
Vague claimsParaphrasing instead of citing specific dataExtract concrete numbers from abstracts or full text
Missing contextSkipping Phase 1 scopingAlways define scope before searching
Uncritical synthesisNot considering methods/limitationsWhen abstract lacks detail, retrieve full text for essential sources
Citation errorsPost-hoc citation insertionCite while writing, not after
Narrow perspectiveToo few sourcesSearch broadly in Phase 2
Hidden abstract-only useNot disclosing which sources were abstract-onlyAlways include Source Transparency disclosure
Wrong PDF in ZoteroUploading without content verificationVerify (title+authors+content) → upload → validate

Paywalled Sources

Many high-impact journals (Lancet, NEJM, Nature, JAMA, Cell, Science) restrict full-text access to subscribers. This is a normal part of scientific literature, not a workflow failure.

What you CAN do with abstract-only sources (paywalled or otherwise):

  • Cite specific quantitative results reported in the abstract (e.g., "HR 0.67, 95% CI 0.54-0.82")
  • Reference the study design and sample size stated in the abstract
  • Cite the main conclusion as stated by the authors

What you CANNOT do:

  • Describe methodology details not in the abstract
  • Discuss subgroup analyses, secondary endpoints, or limitations not mentioned in the abstract
  • Use the source as a primary reference for interpretive or mechanistic claims

In all cases: include abstract-only sources in the Source Transparency disclosure.


Dependencies

This skill has no hard dependencies. It works with whatever tools are available. For optimal results, the following tools enhance the workflow:

  • PubMed MCP server: Structured biomedical literature search
  • Zotero MCP server: Reference management and citation injection
  • web_search / web_fetch: General source discovery and full-text retrieval
  • docx skill: Professional document generation
  • pdf skill: PDF reading and generation