# Humanities Writing Companion is a SKILL.md, not a library, and it claims twelve modes > An agent skill in the open SKILL.md format that walks a long-form humanities paper from research question through four-layer critique, adversarial review and AI-use disclosure. Its citation tooling finds candidates and stops there on purpose, and the repository description counts eleven modes where the README counts twelve. **tizzy916/humanities-writing-companion** — End-to-end humanities writing assistant — an Agent Skill (open SKILL.md format). 11 modes from Socratic research-question sharpening through AI-use disclosure. Bilingual (EN/Chinese), discipline-aware (literature/history/philosophy/art/religion/linguistics + cross-disciplinary). Four-layer critique, calibratable devils advocate, voice preservation. - Repository: https://github.com/tizzy916/humanities-writing-companion - Website: https://doi.org/10.5281/zenodo.20280773 - Stars: 431 · Forks: 46 - Language: Python - License: NOASSERTION - Published: 2026-09-20 · Updated: 2026-09-20 · Language: en - Canonical page: https://hysenlabs.com/projects/tizzy916-humanities-writing-companion ## The deliverable is a SKILL.md file, and Python is only the scripts beside it Humanities Writing Companion ships as an agent skill in the open Agent Skills format, with the specification at agentskills.io. The substance of it is two files at the repository root, `SKILL.md` and a Chinese counterpart `SKILL.zh.md`, and the format is meant to load into Claude Code, the Claude Agent SDK, or any agent that supports the format. GitHub classifies the repository's primary language as Python, which is misleading about where the work is. Python belongs to `scripts/`, which holds the citation toolchain. The behaviour lives in the Markdown. The target reader is stated precisely: a scholar whose primary deliverable is a long-form argumentative text, across history, philosophy, literature, cultural studies, art history, religious studies, classics, intellectual history and science studies. The positioning is end to end rather than point solution, laid out as one line: ``` research question → literature map → planning → drafting → revision → adversarial review → AI-trace cleanup → blind-reading check → AI-use disclosure ``` The README is equally clear about what it is not: not a polishing tool, not a citation manager, not a research pipeline. ## Twelve modes, and a repository description that says eleven The mode table is organised by stage rather than by feature, and the letters run H through L with A and B in the middle, which suggests the set grew by accretion. Pre-writing holds Mode H for research-question sharpening, Mode I for literature mapping and Mode J for plan-only outlining. Drafting splits into Mode C for conception and new content, and Mode A for paragraph dialogue. Review carries Mode B, the four-layer chapter critique, and Mode D, the devil's advocate with a calibratable one to five intensity and a methodology-focus variant. Revision holds Mode E for writing-bottleneck assistance and Mode F for draft revision with a revision-coach sub-mode. Before submission sit Mode G for blind reading and Mode K for AI-use disclosure. After review, Mode L handles integrating defence or external-review comments through a revision-dossier system. Counting the letters gives twelve, while the repository's own description says eleven modes, so the summary undercounts the table it is describing. Anyone citing the mode count should count the table. Mode L is the one that closes the loop, and it is the only mode in the post-review column rather than the pre-submission one, which suggests the author expects to be using this against referee reports rather than only before submitting. ## Four critique layers, and a rule that decides which one answers first Mode B does not polish along a single axis. It walks four layers, each with its own question: ``` 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?" ``` The order is the argument. Foundation comes before structure, structure before the paragraph, and the sentence layer sits last because a well-phrased sentence in a broken argument is still a broken argument. What is more interesting is the scope-aware priority rule that sits on top. In a full review the skill starts with consequential argument issues. But if you asked for a local edit, it completes that edit when the edit does not depend on resolving the larger issues first. In practice that means a request to tighten one paragraph gets a tightened paragraph rather than a lecture about the thesis, unless the paragraph cannot be tightened without settling something upstream. That rule is the difference between a critic and a blocker, and most single-axis writing tools get it wrong. ## Reader D gets confused on purpose, and the reviewer retracts when it loses Mode D runs four voices rather than one, and three of them are reviewers while the fourth is not: ``` Reviewer A · Theoretically demanding Reviewer B · Historically empirical Reviewer C · Methodologically skeptical Reader D · Well-intentioned but confused ``` Reader D is the distinctive one. The design note is that places where a friendly reader gets confused are weak points in the argument, which turns politeness into a diagnostic instrument rather than a compliment. The anti-sycophancy half is stated as a pair of rules that pull in opposite directions. The reviewer must not concede merely under pressure. But it must retract a challenge when a single decisive source passage or valid argument defeats it, and reviewer intensity never licenses ignoring a correction. A critic that always wins is as useless as one that always agrees, and the project treats the second failure as the more common one. Critique strategy also varies by chapter type. Historical narrative is checked for anachronism, counterfactual stress and source handling. Philosophical argument is checked for conceptual derivation. The list is cut off mid-entry, so the remaining chapter types and their dimensions are named in the contents but not written out here. ## Voice is treated as evidence about the author, not as decoration The first of three commitments the project takes seriously is that voice preservation carries epistemic weight. The reasoning is specific rather than sentimental: a writer's voice signals which intellectual tradition they write from, which interlocutors they take seriously, and which moves are theirs and which are borrowed. A paper polished into standard academic English loses that signal. So the target is stated as writing more like yourself, not less, and the profile goes past sentence preferences into four further dimensions: argumentative rhythm, described as linear, spiral or variable-tempo; scholarly posture, as critical inheritance or dialogical advance; the rhetorical function of a citation, which can be an authority anchor, a critical target, a dialogue interface, a narrative move or a conceptual tool; and an AI-trace checklist of eight categories of unexamined pattern, held apart from both AI cliché and scholarly cliché. The second commitment is that argument and prose are inseparable. The contrast drawn is with empirical work, where a good experiment can be ruined by bad writing, whereas in the humanities a slack sentence, a vague concept or an unwarranted transition is itself an argumentative failure. That is why the skill operates at the level of argument through prose instead of on sentences in isolation. ## The citation toolchain finds candidates and says outright that it stops there The one part of this project that is Python is `scripts/`, and its scope is deliberately narrow. It does consistency checking across a manuscript, converts between citation formats including Chicago, MLA, APA and GB7714, and looks up metadata candidates through Crossref and OpenAlex. The parenthetical in the README matters more than the feature list: this is not source-text or claim verification. The tool proposes a candidate record for a citation; it does not read the cited work and confirm the source supports the sentence. For a discipline where the difference between a source that supports a claim and one that merely looks right is the whole ballgame, that is a load-bearing admission rather than a caveat. In an agent-capable environment such as Claude Code, Modes B and D can fan out to parallel reviewer agents, and pending claims can be checked against sources with provenance, support and limitations recorded in separate fields. Recording those three separately is the part that matters for a later audit, because it stops a partly supported claim from being filed as a settled one. ## The DOI in the repository metadata is the one the project already fixed The Zenodo concept DOI carries a version mismatch worth checking before you cite this work. The README badge and the skill header both point at `10.5281/zenodo.20280772`. The repository's homepage field still holds `10.5281/zenodo.20280773`, which is the pre-fix value, and the v5.0.2 release title reads `Fix Concept DOI (20280773 → 20280772)`. So the correction shipped in July 2026, and one field out of the repository did not move with it. If you are citing the concept DOI, take the number from the README badge. Licensing is the other field to read carefully. A `LICENSE` file is present at the root, alongside `CITATION.cff` and `.zenodo.json`, but the repository's licence field reports NOASSERTION rather than an identifier, which is what GitHub reports when it cannot classify the file. The file itself is the authority here, not the sidebar. The root also carries a `CHANGELOG.md`, `CONTRIBUTING.md` with a Chinese counterpart, both READMEs and both skill files, plus `docs/` and `references/`. Version 5.1.0 shipped on 2026-09-22 with scoped workflows and citation fixes, and the last push to the repository was on the same date. ## Install, quick start and the tool comparison are named in the contents and absent from the text The table of contents lists fifteen sections. The visible text covers positioning, what the skill takes seriously, a typical interaction, and the first half of core features. Everything after that stops mid-entry on the discipline-specific dimensions. That cut takes out Install, Quick start with three typical scenarios, Comparison with adjacent tools, Project structure, Design philosophy and Citation. For a project whose value depends on how it is loaded into an agent, losing the Install section is the expensive part of the truncation, since the format requirement is all that survives: an open SKILL.md at agentskills.io, loaded into Claude Code, the Claude Agent SDK or any agent supporting the format. The comparison section is a similar loss, and the project says enough about itself to substitute for part of it. Not a polishing tool, not a citation manager, not a research pipeline. And the citation toolchain's own boundary, metadata candidates rather than verification, tells you where this one stops relative to a reference manager. What the visible text does show is a worked interaction, in which a reader asks about extending Foucault's panopticism to digital platforms and gets back a foundation-level objection that the metaphor imports architecture's subject-position where platforms produce a different subject entirely, plus two named ways out. ## Conclusion Adopt this skill if you write long-form humanities prose and want a critic that treats a slack sentence as an argumentative failure rather than a style note, since that is the premise the whole mode set is built on. Do not adopt it as a citation checker, because the project says its own tooling stops at metadata candidates and format conversion, and do not expect the README to tell you how to install it, since the contents promise an Install section that is not in the file. Before the first run, read SKILL.md rather than the README, and confirm the two DOIs match what you cite, because the repository metadata field still carries 20280773 while the README badge and the v5.0.2 release both point at 20280772. ## FAQ ### What is the Humanities Writing Companion and how is it loaded? It is an agent skill written in the open Agent Skills SKILL.md format, kept as SKILL.md and SKILL.zh.md at the repository root. It runs in Claude Code, the Claude Agent SDK, or any agent that supports the format described at agentskills.io. ### How many modes does the Humanities Writing Companion have? The README table lists twelve: H, I and J before writing, C and A for drafting, B and D for review, E and F for revision, G and K before submission, and L after review. The repository description says eleven. ### Does the Humanities Writing Companion verify that my citations are correct? No. Its citation toolchain does consistency checking, format conversion between Chicago, MLA, APA and GB7714, and metadata candidate lookup through Crossref and OpenAlex. The README states this is not source-text or claim verification. ### What licence is the Humanities Writing Companion under? A LICENSE file sits at the repository root, but the repository's licence field reports NOASSERTION, so read the file rather than the sidebar. A CITATION.cff and a .zenodo.json are also in the root. ## Sources - [Issues](https://github.com/tizzy916/humanities-writing-companion/issues) - [Project website](https://doi.org/10.5281/zenodo.20280773) - [README](https://github.com/tizzy916/humanities-writing-companion/blob/main/README.md) - [Releases](https://github.com/tizzy916/humanities-writing-companion/releases) - [tizzy916/humanities-writing-companion on GitHub](https://github.com/tizzy916/humanities-writing-companion) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/tizzy916-humanities-writing-companion