Paper Writing Skills Index

Paper Writing Skills

Mode

Two modes control which files and gates are loaded. Set mode in triage.md during Phase 0.

If any submission trigger word appears in the user request, set mode: submission regardless of scenario. Otherwise default to report.

Routing + Sequential Pipeline

Scenario Routing

Family Index

Scenario Format Index

HTML output for any format must satisfy the Editable HTML Contract below.

Required Outputs

• {workdir}/triage.md
• {workdir}/draft/paper.md (Markdown source of truth)
• {workdir}/report/<slug>_preview.html
• {workdir}/quality/claim_evidence_report.md
• {workdir}/quality/reviewer_report.md — submissions, grants, rebuttals, high-risk tasks only

Rendering Templates

Standard rendering flow

1. Read report_standard.md for the HTML+CSS template.
2. Convert draft/paper.md to HTML, wrapping major sections per the Editable HTML Contract below.
3. Apply the selected theme. Write {workdir}/report/<slug>_preview.html.

Academic rendering flow (LaTeX → PDF)

1. Read the LaTeX template (latex_cn.md or latex_en.md based on language).
2. Convert draft/paper.md to LaTeX, fill the template, write <slug>.tex.
3. Read COMPILE.md — probe available engines before compiling.
4. Also produce an HTML preview via report_academic.md.

---

Evidence Layer

Evidence skills decide where facts come from. Claims may not be invented or grounded in model memory; they must trace back to a registered piece of evidence.

Default pipeline: paper_fetch → evidence_summary → evidence_registry → draft → citation grounding

Evidence Registry — claim_evidence_map.md

Evidence types: citation, figure, table, experimental_data, statistical_result, user_material, missing. Claims with missing evidence cannot appear as firm conclusions. If support is partial, narrow the wording.

Evidence Summary — output table

Preserve locator detail for later attribution. Separate the source's statement from the agent's interpretation. If evidence does not answer the query, say so; do not force fit.

Context-Bound Answering

Answer only from provided context. Cite context keys, evidence IDs, pages, or material IDs. If context is insufficient, write I cannot answer from the provided context and list what is missing. Do not fall back to model memory.

---

Quality Gates

Gates produce reports and risk flags — guardrails, not gatekeepers. Flag issues and suggest fixes; do not silently rewrite a draft to hide problems. Aim for ≥80% compliance; prioritize critical issues (missing data, unsupported claims, fabricated citations) over cosmetic formatting.

Run order: manuscript coverage → format lint → reproducibility → claim/evidence → reporting guideline → reviewer simulation → HTML editability → response consistency.

Manuscript Coverage Check

Core sections (required): Title, Abstract, Introduction, Methods, Results, Discussion, References. Optional (journal-dependent): Keywords, Conclusion, Acknowledgements, Author Contributions, Competing Interests, Data/Code Availability, Supplementary Materials.

Output: ## Manuscript Coverage Report with completeness %, present/complete list, missing/incomplete list, recommendation. Gate: ≥90% excellent / 80-89% good / <80% major sections missing.

Format Lint

Check: section structure complete; figure/table numbering sequential with captions; references numbered, no gaps, consistent format; word/page limits met; supplementary files referenced and numbered.

Output: ## Format Lint Report with compliance %, correct categories, issues (severity + location + action), word count vs limit, recommendation.

Common limits (always verify against journal guidelines):

Gate: ≥90% excellent / 80-89% fix flagged / <80% major problems.

Reproducibility Check

Audit categories (each must be present and concrete): software & tools (names, versions, OS, libraries); parameters & settings (hyperparameters, seeds, hardware); data (source, version, splits, availability); code (repo URL, license, dependencies, reproducibility script).

Output: ## Reproducibility Check Report with compliance %, sufficient detail list, missing/vague list (current text + fix), availability statement templates, recommendation. Gate: ≥90% excellent / 80-89% good / 70-79% fair / <70% major revision.

---

Editable HTML Contract

All HTML outputs must be standalone working files, not app-only previews.

<section class="editable-block" contenteditable="true"
  data-block-id="abstract-001" data-section="abstract"
  data-source="draft/paper.md" data-format-role="summary">
  <h2>Abstract</h2><p>…</p>
</section>

Required: single file with embedded CSS; semantic tags (section, figure, figcaption, table, caption); @page or @media print; no frontend runtime required; no absolute positioning for body layout.

Validation (run after every HTML generation): file opens standalone; CSS embedded; main text is semantic HTML not an image; major blocks have contenteditable="true" and all four data-* attributes; print CSS present; figures/tables use semantic tags. If any criterion fails, fix the renderer or source Markdown — do not relax the contract.

---

Non-Negotiable Rules

• Do not write unsupported claims. Mark missing evidence and downgrade or remove the claim.
• Do not invent citations, DOIs, accession IDs, page numbers, data repositories, reviewer changes, or experimental results.
• Do not let a search result become evidence until it is summarized, attributed, and bound to a specific claim in the Evidence Registry above.
• Do not use Sci-Hub or any access-control bypass. Open PDF fetching may use only legal OA routes in paper_fetch.md.
• Do not output screenshot-like reports. Main text must remain semantic and editable HTML per the Editable HTML Contract above.
• Keep each SKILL.md below 500 lines.

Sources

• Anthropic skill design — https://github.com/anthropics/skills/blob/main/skills/skill-creator/SKILL.md
• Anthropic doc coauthoring — https://github.com/anthropics/skills/blob/main/skills/doc-coauthoring/SKILL.md
• Trail of Bits workflow skill design — https://github.com/trailofbits/skills/blob/main/plugins/workflow-skill-design/skills/designing-workflow-skills/SKILL.md
• PaperQA evidence RAG — https://github.com/Future-House/paper-qa
• OpenScholar attribution — https://github.com/akariasai/openscholar
• Nature-style response/figure/citation/data skills — https://github.com/Yuan1z0825/nature-skills
• DeepScientist outline/review/rebuttal/writer skills — https://github.com/ResearAI/DeepScientist
• Research-Paper-Writing-Skills (Master-cai, MIT) — abstract / introduction / method / results / discussion section guides
• AI-Scientist (SakanaAI, MIT) — NeurIPS-style reviewer rubric
• EQUATOR Network — reporting guideline checklists
• tw93/Kami (MIT) — academic theme