Academic Writing Skills for Claude Code

中文版

This collection of skills grew out of my day-to-day paper-writing workflow and
has been iteratively refined over time. It may still have shortcomings or rough
edges; if needed, please fork it and adapt it yourself.

I have recently been writing my dissertation and revising English papers,
so I will keep improving latex-thesis-zh and latex-paper-en based on
actual usage.
Note: paper-audit review reports are for reference only; please verify and
evaluate them yourself.

Post-writing polish and validation for academic papers: format checks,
bibliography search and verification, grammar analysis, de-AI editing, and
experiment narrative review. Focused on improving existing drafts, not writing
papers from scratch.

Recommended models (★ is most recommended): ★ Gemini 4 Carbon / Gemini 4 Argon · Claude Opus 5.5

Google's models have a good writing style, but they may have some
hallucinations when writing science and engineering papers. Review carefully.

GPT models do not write like a person, and the defensive hedging is serious.
If you use one, watch both. After GPT has revised the text, scan it again
with another model, and read it yourself.

Recommended platforms: Claude Code · Pi · OMP (Oh My Pi)

Install

Install the repository with skills:

npx skills add bahayonghang/academic-writing-skills

Then open an applicable tool in your manuscript project and ask for the task in
natural language. The root README is only a routing guide; the authoritative
usage details live in each SKILL.md file and in the docs site.

The LaTeX loaders reject includes outside the project root. For installation
alerts and E-INCLUDE-BOUNDARY, see the installation guide.

Choose A Skill

Opt-in checks only report local candidates: check_consistency.py --governance requires --custom-terms, --abbreviation-style is independent, and check_style_zh.py --degree-wording is off by default; without these flags the previous output stays unchanged.
College number, equation, table-body, and Chinese-caption checks use --school yanshan-ee-2025 on check_style_zh.py, check_format.py, check_tables.py, and check_references.py. The default and --school generic add no candidates, and there is no bare yanshan alias.
Citation placement, repeated-citation pages, college bibliography prompts, and review progression density run only behind explicit flags: check_references.py --author-cite and --repeat-cite are independent and may combine with --school; verify_bib.py --college-details is legal only with --standard gb7714 or gb7714-2025; analyze_literature.py --progression-density may combine with --section and is mutually exclusive with --intro-citations. Without these flags the previous output stays unchanged.
Same-chapter table, body, and chapter-summary final values are compared only by analyze_experiment.py --cross-surface. --section may narrow the chapter. --cross-surface-terms FILE is valid only with that switch and replaces only the metric or evaluation-set list. Candidates are [Script], Info/P3, and Meaning-Check: NEEDS-LLM, with a local position and no corrected number. Without --cross-surface the previous output stays unchanged, and the existing --results-analysis codes stay independent.
The graduate-school checklist is --template yanshan. The 2025 college checklist is --template yanshan-ee-2025. The two coexist. 111 statuses are not 111 compliant items. A partial checker does not PASS a compound item, and MODULE or NEEDS-LLM still needs human review.
Method-expression labels, a weakness written as an advantage, a pronoun left without an antecedent after a deleted preview, abstract quotation marks, and formula symbols in a title or chapter-arrangement line are LLM-only readings on the existing logic, claim-forward, abstract, and structure modules. They add no script code, no threshold, and no new module.
The claim-forward guides now add four LLM-only dispositions (pre-emptive rebuttal shell, concession that dismisses a confound, over-strong claim with a vague concession, empty modesty), and the rewrite contract in latex-paper-en, latex-thesis-zh, and typst-paper now forbids both raising and lowering claim strength, with a new LLM-only understatement flag. They add no script code, no threshold, and no new module.

Use the format-specific writing skills when you want source edits or polish. Use
paper-audit when you want reviewer-style diagnosis without rewriting the
source. Use bib-search-citation when the target is the bibliography library
itself.

Typical Prompts

Compile my English LaTeX paper with latexmk and explain the first blocking error.

Check the introduction for logic gaps, citation stacking, and AI-like phrasing.

Run a paper-audit gate on main.tex and separate blockers from polish issues.

Search references.bib for recent Mamba forecasting papers with code and return LaTeX and Typst citation snippets.

Align-check this cover letter against main.tex and report unsupported claims only.

Safety And Outputs

• The skills are for improving and validating existing academic material. They
  should not invent experiments, citations, policies, or unsupported claims.
• Citation keys, DOI, arXiv IDs, URLs, and local .bib matches are provenance
  fields, not proof that a paper supports a manuscript claim.
• Online checks are optional. When current venue rules or external metadata
  matter, verify them from the original source before treating them as binding.
• Source-editing suggestions should preserve LaTeX and Typst syntax and mark
  required evidence as pending instead of filling it in.
• Audit and helper-script outputs may be JSON, Markdown reports, or diff-comment
  style findings with severity and priority.

Example finding shape:

% <MODULE> (Line <N>) [Severity: Critical|Major|Minor] [Priority: P0|P1|P2]: <Issue summary>
% Before: <original text>
% After:  <suggested text>
% Rationale: <brief explanation>
% [PENDING VERIFICATION]: <if evidence or metrics are required>

Requirements

• Python 3.10+
• uv for running the bundled Python helpers
• TeX Live or MiKTeX with latexmk and chktex for LaTeX workflows
• XeLaTeX plus CJK fonts for Chinese LaTeX documents
• Typst CLI for Typst workflows
• pymupdf (PyMuPDF) for PDF audit workflows; optional pymupdf4llm for enhanced extraction (both lazily imported)
• Node.js and npm or just only when building the docs site locally

Repository Layout

academic-writing-skills/
├── academic-writing-skills/
│   └── <skill>/
│       ├── SKILL.md          # Skill entrypoint, triggers, routing, contracts
│       ├── scripts/          # Optional executable helpers
│       ├── references/       # Optional source-of-truth guidance
│       ├── examples/         # Optional example prompts or workflows
│       ├── templates/        # Optional output templates
│       ├── evals/            # Optional evaluation cases
│       └── agents/           # Optional agent metadata
├── docs/                     # Documentation site
├── tests/                    # Pytest coverage for contracts and helpers
├── ref/                      # Supporting reference repositories or material
├── .trellis/                 # Project workflow and guidance
├── README.md
└── README_CN.md

Do not rely on the root README for module internals. Open the relevant
SKILL.md, references/, and docs page when a workflow needs exact routing,
flags, or output contracts.

Documentation

Full documentation is in docs/ and docs/zh/.

Run locally:

just docs

Build the static site:

just doc-build

Contributing

Issues and pull requests are welcome. Keep changes scoped to the relevant skill,
update tests or docs when behavior changes, and run just ci before submitting
when feasible.

License

PolyForm Noncommercial License 1.0.0.
Academic research, teaching, personal study, and other noncommercial use are permitted.
Commercial use is not permitted.

Third-party material keeps the license named in its own source note.