Humanities Writing Companion · 人文学科写作伙伴

A Claude Code / Claude Agent SDK skill for humanities scholars whose primary deliverable is a long-form argumentative text — history, philosophy, literature, cultural studies, art history, religious studies, classics.

📖 Wiki · 中文版 README · Skill source · English · Skill source · 中文

Positioning

End-to-end writing assistant for humanities scholars — covering the full lifecycle of a humanities paper from research question to submission disclosure:

research question → literature map → planning → drafting → revision →
adversarial review → AI-trace cleanup → blind-reading check → AI-use disclosure

Built for fields where prose IS the argument — history, philosophy, literature, cultural studies, art history, religious studies, classics, intellectual history, science studies, and adjacent humanities-aligned fields.

Not a polishing tool. Not a citation manager. Not a research pipeline. A thinking partner that stays with you across the whole arc.

11 modes covering the writing lifecycle

Plus a citation toolchain (scripts/): consistency checking, format conversion (Chicago / MLA / APA / GB7714), and reference verification against Crossref.

What this skill takes seriously

Three things generic AI writing tools do not handle for humanities:

1. Voice preservation is a core scholarly value, not stylistic decoration. In humanities, the author's voice carries epistemic weight: it signals which intellectual tradition the author writes from, which interlocutors they take seriously, which moves are theirs and which are borrowed. A paper polished into "standard academic English" loses this signal. This skill helps you write more like yourself, not less.

2. Argument and prose are inseparable. In empirical research, you can have a perfect experiment ruined by bad writing. In humanities, the writing IS the argument — a slack sentence, a vague concept, an unwarranted transition is an argumentative failure. This skill works at the level of argument-through-prose.

3. The reviewer is real and adversarial. Humanities reviewers are not gentle. A theoretical concept will be tested for sharpness; a historical claim will be tested for evidence; a philosophical argument will be tested for the strongest counter. This skill simulates that adversary internally so the paper meets it before submission.

A typical interaction

You: 帮我看看这段。我在论福柯的全景敞视主义如何延伸到数字平台。

[skill reads the paragraph + your style profile]

Companion:
  I notice three things, in order of priority.

  🔴 Foundation: you're using "全景敞视" as a metaphor for platform 
     surveillance, but Foucault's original concept turns on architecture 
     producing a specific subject-position. Platforms produce a different 
     subject (data subject, not body in a cell). You may want to either 
     (a) name the analogy as analogy and discuss the difference, or 
     (b) borrow Zuboff's "instrumentarian power" which fits the data case 
     more cleanly. Which direction do you want to go?

  🟡 Argument: paragraph 2 jumps from "platforms collect data" to "this 
     constitutes surveillance" without the middle step about visibility-
     as-power. A friendly reader will not follow.

  🟢 Voice: your style profile has "first-person with em-dash insertion." 
     Paragraph 3 starts with "本文认为…" which reads like cliché. Want to 
     restore the first-person?

You: 走方向 (b). 改吧。

[skill produces revision; you keep the version that sounds like you]

This is what "thinking partner, not polishing tool" means in practice.

Core features

Four-layer critique (not single-axis polishing)

Layer 1 · Foundation   "Does this paper hold up scholarly?"
Layer 2 · Structure    "How is the argument unfolding?"
Layer 3 · Paragraph    "What is this paragraph doing?"
Layer 4 · Sentence     "Is this sentence right? Well-said?"

Strict top-down rule: do not exert effort at lower layers while upper layers are unresolved.

Devil's advocate with anti-sycophancy

Simulates three reviewers + one well-intentioned reader:

• Reviewer A · Theoretically demanding
• Reviewer B · Historically empirical
• Reviewer C · Methodologically skeptical
• Reader D · Well-intentioned but confused (distinctive design: places where a friendly reader gets confused are weak points in the argument)

Anti-sycophancy hard rule: when the author pushes back on a challenge, the AI must see at least 2 of 5 substantive conditions met before conceding — prevents premature softening under emotional pressure.

Deep voice learning and preservation ("My hand writes my voice")

Not just sentence preferences. Also:

• Argumentative rhythm (linear / spiral / variable-tempo)
• Scholarly posture (critical inheritance / dialogical advance)
• Rhetorical function of citation (authority anchor / critical target / dialogue interface / narrative / conceptual tool)
• AI-trace checklist (8 categories of unexamined patterns ≠ AI cliché ≠ scholarly cliché)

Discipline-specific dimensions

Different chapter types get different critique strategies:

• Historical narrative · anachronism, counterfactual stress, source handling
• Philosophical argument · conceptual derivation, cross-theoretical transplantation, strongest objection
• Literary criticism · close reading vs. interpretation, genre awareness, form and meaning
• Cultural studies · power-knowledge framing, positionality, generalization range
• Art history · description vs. interpretation, provenance, reception
• Religious studies / classics · source-language rigor, tradition awareness, insider-outsider position
• Case analysis (cross-disciplinary)

ADHD-friendly interaction

• Feedback batched 3-5 per round
• Quick wins first
• Topic-jump support (a jump may be an insight signal)
• Pomodoro-friendly task units
• Reorientation summaries every 4-5 turns

Reflexive writing support (distinctive module)

If your research itself involves human-AI collaboration (autoethnography of AI-assisted writing), the skill provides six categories of "reflexive moments":

🔄 Direction change · 🚫 Refusal · 🎭 Voice conflict · 🔧 Tool dependency · 💡 Unexpected insight · 🤖 AI-trace awareness

Scholarly basis (citable in your paper):

• Christou (2026). Reconfiguring Reflexivity in the Era of AI. Qualitative Inquiry.
• Wiles (2025). Recursive Cognition in Practice. International Journal of Qualitative Methods.
• Panke (2025). How Can (A)I Research This? An Autoethnographic Exploration of Generative AI.

Engineering infrastructure

Borrowed from software engineering, in service of humanities writing:

• Version management: minor version = commit / major = release / _drafts/ = feature branch
• Revision log: every change recorded with diff + reason (like git commit message)
• Feedback reports: Blocker / Major / Minor / Question tiers (from code review)
• Systematic verification checklists: argument completeness / concept consistency / citation completeness / style consistency
• [VERIFY] / [待核对] hard marker: anti-citation-hallucination — cannot enter submission version

Seven work modes (not one)

• Mode A — paragraph-level dialogue
• Mode B — chapter-level review
• Mode C — conception → new content writing (with collaborative drafting protocol)
• Mode D — devil's advocate
• Mode E — writing bottleneck assistance (5 unblocking strategies)
• Mode F — draft revision (two-version comparison, anti-AI-cliché)
• Mode G — blind reading (mechanical promise-delivery check)

Engineering helper scripts

scripts/ provides three zero-dependency tools:

Supported humanities disciplines

This skill organizes humanities scholarship in a three-layer architecture, so the discipline-routing system can match where the author actually works — not just to one of a flat list of seven slots. Authors declare their discipline at onboarding (or the skill infers from the draft); routing then loads the matching layer.

L1 · Six humanities main disciplines

These are the canonical L1 humanities disciplines. Each carries a core set of methodological concerns generic AI writing tools miss.

L2 · Common subfields (not exhaustive)

Subfields inherit all the methodological concerns of their parent L1, plus any specific constraints the author declares at onboarding. Examples — many more possible:

If your subfield isn't listed, declare it at onboarding — it inherits from its parent L1 automatically.

L3 · Cross-disciplinary fields (explicit multi-inheritance)

These are humanities fields that explicitly draw from multiple L1s. The skill loads the methodological concerns of all parent L1s plus the L3-specific overlay.

Humanities-adjacent fields (welcome with scope note)

Some fields are formally classified as social science but include strong humanities-style sub-traditions (where prose IS the argument). This skill welcomes such work:

Inheritance: communication studies (humanities-style) ← Media studies + Philosophy + Cultural studies. Educational research (humanities-style) ← History + Philosophy + Cultural studies.

Fallback protocol — if you still don't see your field

Tell the skill at onboarding:

1. Object of study: text / past event / concept / phenomenon / artifact / practice / language structure / …
2. Primary method: close reading / archival research / argument analysis / ethnography / comparison / philological work / …

The skill will infer your closest L1 + any relevant L3 overlays, write it to _writing-config/discipline.md, and proceed. You can refine the declaration anytime.

The full methodology rubric for each L1 / L3 entry lives in SKILL.md § Humanities Discipline-Specific Dimensions — this README table is the surface map; the skill carries the full rubric.

Showcase: Before / After

A real example of the four-layer critique in action. Author's draft paragraph (English translation of a Chinese-language thesis chapter on Stiegler):

Before. Stiegler's pharmakon concept shows that technology is both poison and remedy. AI is therefore a pharmakon as well. We must be careful in our use of AI, lest we fall into its negative effects. As Stiegler said, technology is a double-edged sword.

The skill's response, top-down, four layers:

🔴 Layer 1 · Foundation — Argument standing

  The paragraph claims pharmakon = "double-edged sword." Stiegler explicitly
  rejects this equivalence (WMLWL, p. 12): pharmakon is curative AND
  destructive AT THE SAME TIME, not alternately, and the destructive force
  is proportional to the curative one. "Double-edged sword" implies a
  user-choice framing; pharmakon implies a structural condition that
  precedes user choice. This is not nuance — it is the whole point of
  borrowing the concept from Derrida.

  Verdict: Foundation broken. Do not proceed to lower layers until fixed.

🟡 Layer 2 · Structure — Argument unfolding (deferred until L1 fixed)

🟢 Layer 3 · Paragraph (deferred)

⚪ Layer 4 · Sentence (deferred)

After the author rewrites the foundation, the same paragraph might become:

After. In Stiegler's framework — extended from Derrida's reading of Plato's Phaedrus — pharmakon names a structural condition rather than a moral choice: a technology's curative force is inseparable from and proportional to its destructive force (Stiegler, What Makes Life Worth Living, 2013, p. 12). For LLMs, this means the question is not "are we careful enough in our use?" — that frame presupposes a user fully outside the pharmakon. The question is: in what historical-organological configuration does the pharmakon's destructive face become structurally dominant? I argue, following Stiegler's reading of digital tertiary retention in Automatic Society (2017), that …

What changed: a clichéd "double-edged sword" framing replaced by Stiegler's actual conceptual move, a citation anchored at a verifiable page, and a forward-pointing thesis the next paragraph can develop. The skill did not write the rewrite — it identified that the foundation was wrong, named why, and refused to do sentence-level work until the foundation was repaired.

This is what "thinking partner, not polishing tool" means in practice.

Install

As a Claude Code skill

git clone https://github.com/tizzy916/claude-skill-humanities-writing-companion.git \
  ~/.claude/skills/humanities-writing-companion

chmod +x ~/.claude/skills/humanities-writing-companion/scripts/*.sh

Or as a project-level skill (vault / project only):

git clone https://github.com/tizzy916/claude-skill-humanities-writing-companion.git \
  ./.claude/skills/humanities-writing-companion

Claude Code loading

Claude Code auto-scans ~/.claude/skills/ and ./.claude/skills/ on startup. After install, say "I'm working on a humanities paper" or any of the trigger phrases below.

Claude Agent SDK

SKILL.md can be loaded into your system prompt directly. The skill is plain text — no runtime dependencies.

Trigger phrases

English: "paper," "essay," "chapter," "dissertation," "argument," "thesis," "revise," "voice," "review my section," "stuck on writing," "devil's advocate," "reviewer attack"

Chinese: 论文 · 写作 · 润色 · 改论文 · 帮我看看这一章 · 我手写我口 · 这个论证有没有问题 · 我写不下去了 · 审稿人会怎么攻击

Even casual mentions trigger: "take a look at this paragraph" · 帮我看看这段话

Quick start · three typical scenarios

Scenario 1: new paper, first use

Say to Claude: "I want to write a paper on X."

The skill enters onboarding: confirms citation format, target reader, existing writing samples, and initializes the project folder structure.

Scenario 2: revising an existing chapter

"Help me read this chapter"      → Mode B (chapter review) → 4-tier feedback report
"Help me revise this paragraph"  → Mode A (paragraph dialogue) → diagnose + suggest + reason
"I'm stuck"                      → Mode E (bottleneck) → 5 unblocking strategies

Scenario 3: fighting AI cliché (two-version comparison)

If your paper has been through AI polishing but you want to restore the original voice:

Mode F · draft revision → compare AI-polished vs. original → keep improvements + restore voice

Comparison with adjacent tools

Project structure

humanities-writing-companion/
├── SKILL.md                          ← Main skill file (EN, ~1400 lines, 11 modes)
├── SKILL.zh.md                       ← Chinese mirror (中文版)
├── references/
│   ├── ai-trace-checklist.md         ← AI-trace scan checklist (currently Chinese; EN translation TODO)
│   ├── project-management.md         ← Project folder + version management
│   └── target-reader-profile-template.md  ← Target reader profile template
├── scripts/                          ← Engineering toolchain (zero deps)
│   ├── README.md                     ← Script usage
│   ├── ai-trace-scan.sh              ← AI cliché scan (zsh)
│   ├── pending-checks.sh             ← Pending marker aggregation (zsh)
│   ├── citation-consistency.py       ← Citation format consistency (Python 3)
│   ├── citation-format-convert.py    ← Chicago/MLA/APA/GB7714 converter (v4.0+)
│   └── citation-verify.py            ← Crossref-based citation verification (v4.0+)
├── README.md                         ← This file
├── README.zh.md                      ← 中文 README
├── CHANGELOG.md                      ← Version history
├── LICENSE                           ← CC BY-NC 4.0
└── CITATION.cff                      ← Academic citation metadata

Bilingual status: SKILL.md and README are bilingual (EN + CN). references/ files and scripts/ comments are currently primarily Chinese; English translations are TODO. Both languages of trigger work either way (the description field in SKILL.md handles both).

Design philosophy

"My hand writes my voice"

Academic rigor and personal expression are not opposites. "Standard academic prose" usually means the death of individuality. The skill helps the author speak in their own voice rather than pressing their words into a prefabricated mold.

Thought first, format second

Revision priority:

1. Force of the argument
2. Precision of concepts
3. Effectiveness of structure
4. Quality of expression
5. Format compliance

Always top-down. Do not fuss with commas in a paragraph whose underlying argument is broken.

Engineering rigor, humanistic expression

Borrows software engineering best practices (version management, unit tests, code review) in service of humanities writing. Engineering rigor does NOT mean turning the paper into code — it means every revision is traceable, argument quality is verifiable, the writing process is resumable, and problems are processed in layers.

Citation

If your research uses this skill, please cite it in the methodology section.

BibTeX:

@software{shen_humanities_writing_companion_2026,
  author       = {Shen, Cong},
  title        = {Humanities Writing Companion: A Claude Skill for Voice-Preserving Humanities Academic Writing},
  year         = {2026},
  publisher    = {Zenodo},
  version      = {4.1.1},
  doi          = {10.5281/zenodo.20280773},
  url          = {https://doi.org/10.5281/zenodo.20280773}
}

Plain-text attribution (for skill metadata, footers, etc.):

Based on Humanities Writing Companion by Shen Cong
https://github.com/tizzy916/claude-skill-humanities-writing-companion

See CITATION.cff for full machine-readable metadata (GitHub's "Cite this repository" button will use it automatically).

Citing companion tools

If you also use academic-research-skills in the same project, please cite both. ARS attribution format (per CC BY-NC 4.0):

Based on Academic Research Skills by Cheng-I Wu
https://github.com/Imbad0202/academic-research-skills

Contributing

Issues and PRs welcome:

• New work-mode proposals
• Extensions to the AI-trace checklist
• Discipline-specific examples (medieval studies, art conservation, ethnomusicology, etc.)
• Additional citation-format support (APA / Chicago / MLA / GB/T 7714 / journal-specific)
• English translation of references/ files

See CONTRIBUTING.md.

About the Author

Shen Cong — BFA, Experimental Art, Central Academy of Fine Arts (CAFA); MA, History of Science, Tsinghua University (advisor: Hu Yilin); Founder & CEO of Tianyu Vision, a sci-art studio working on scientific visualization, science communication, and sci-art convergence.

This skill came out of writing the author's own MA thesis, Technical Liberalism. He noticed that almost every AI writing tool on the market pulled toward polishing and averaging — whereas humanities scholarship needs the opposite: protecting the author's scholarly voice, stress-testing argumentative rigor, and surviving adversarial peer review. So he built this skill — not to write for him, but to read for him, delivering at each of four layers (basic rigor / argument development / paragraph function / sentence-level expression) the kind of critique a real humanities scholar would actually give.

📮 GitHub @tizzy916 · shencong916@gmail.com · Corrections, collaboration, and conversation welcome.

License

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International) — free for non-commercial use, modification, and distribution. Requires attribution.

⚠️ License change (v3.0.0, 2026-05-19): This project relicensed from MIT to CC BY-NC 4.0. Versions ≤ v2.1.0 remain under MIT and retain their original commercial-use rights for those specific versions. From v3.0.0 onwards, commercial use is prohibited without a separate license.

Commercial Use

This skill is licensed under CC BY-NC 4.0 — non-commercial use only (academic research, teaching, personal projects, open-source derivatives, internal research workflows).

For commercial licensing inquiries — embedding in a paid product, paid consulting or editing services using this skill, commercial SaaS integration, agency use on behalf of paying clients — contact the author for a commercial license:

📮 shencong916@gmail.com (Shen Cong · Tianyu Vision)

The author retains the right to grant commercial licenses on a case-by-case basis. Citing this skill in academic publications is always permitted regardless of license tier.

Acknowledgments

Methodological inspiration and scholarly basis:

• Christou, P. A. (2026). Reconfiguring Reflexivity in the Era of AI. Qualitative Inquiry.
• Wiles, F. (2025). Recursive Cognition in Practice. International Journal of Qualitative Methods.
• Panke, S. (2025). How Can (A)I Research This?.
• Foucault, M. (1984). What is Enlightenment? — "diagnostic of the present" as methodological tradition
• Stiegler, B. (2013). What Makes Life Worth Living: On Pharmacology — "critical pharmacology"

Some design patterns inspired by:

• Imbad0202/academic-research-skills — the upstream pipeline this skill complements; the Concession Threshold pattern in their reviewer module inspired Mode D's "minimum standard before conceding"
• Voice DNA + Audience Profile pattern — inspired the style profile + reader profile pairing
• Thesify Purpose-Check — inspired Mode G blind reading