-
Notifications
You must be signed in to change notification settings - Fork 0
Working Modes
The Humanities Writing Companion is not a single "polish my text" button. It is a set of 11 working modes that map onto the full lifecycle of a humanities paper — from the first vague research interest all the way to the AI-use disclosure you paste into the submission portal.
You never have to memorize the letters. In practice you just say what you need in plain language, and the skill routes you to the right mode. This page exists so you understand what is actually happening under the hood — and so you can call a mode by name when you want fine control.
A few things to keep in mind before the tour:
- You don't pick one mode and stay there. A real writing session hops between modes. You sharpen a question (H), map the literature (I), outline (J), draft a section (C), then a paragraph stalls so you drop into bottleneck mode (E), then you stress-test the argument (D). The skill suggests switches as you go.
- Modes are opinionated on purpose. Some modes will refuse to do what you ask — Mode I won't search literature for you, Mode J won't write a paragraph, Mode H won't bless a topic that isn't yet a question. That refusal is the feature. It's what keeps the scholarship yours.
- Most modes leave a trace. Revisions land in a revision log; adversarial concessions, drafting decisions, and mode history land in an interaction log. That paper trail is what later lets Mode K write an honest AI-use disclosure.
The 11 modes cluster into five phases. Reading this table top to bottom is roughly the arc of a paper.
| Phase | Modes | What you're doing |
|---|---|---|
| Pre-writing |
H · Research-question sharpening I · Literature mapping J · Plan-only outlining |
Turn a vague interest into a sharp, defensible question; organize what you've read into a map of camps and debates; build a section-by-section outline — all before a single paragraph is drafted. |
| Drafting |
C · Conception → new content A · Paragraph-level dialogue |
Move from idea to prose. Mode C is the four-stage flow from conception to integrated draft (AI as midwife, not architect); Mode A is the close, paragraph-by-paragraph workbench. |
| Review |
B · Chapter-level review D · Devil's advocate |
Read the whole thing critically. Mode B is the four-layer holistic critique; Mode D simulates real, adversarial reviewers and refuses to flatter you. |
| Revision |
E · Writing-bottleneck assistance F · Draft revision (two-version compare) |
Get unstuck, and clean up. Mode E breaks writer's block with five tactics; Mode F systematically de-AIs a polished draft while restoring your voice. |
| Pre-submission |
G · Blind reading (promise-delivery) K · AI-use disclosure |
Final gates. Mode G mechanically checks "did you deliver what you promised?"; Mode K writes the AI-use statement the journal now demands. |
Heads-up on the lettering: the letters were assigned roughly in the order the modes were added, not in lifecycle order. That's why pre-writing is H/I/J (added later) while drafting and review are the early A–D. Follow the phases, not the alphabet.
A typical happy path through the lifecycle:
H (sharpen the question)
→ I (map the literature)
→ J (outline)
→ C (draft) ──┐
↕ ├─ E (unstick) when you stall
A (paragraph) ┘
→ B (chapter review)
→ D (devil's advocate) ── F (revise, de-AI)
→ G (blind reading)
→ K (AI-use disclosure) → submit
What it's for. The close-range workbench. You have a specific paragraph — drafted, half-drafted, or just nagging at you — and you want to think about it as an argumentative unit, not just fix its grammar. This is where most day-to-day editing actually happens.
How to trigger it. Paste a paragraph and say something like "take a look at this," "does this paragraph work?", or 帮我看看这段话. Any time you drop a single chunk of prose into the chat, you're in Mode A.
What it uses as input. The paragraph itself, your style profile (so suggestions sound like you), and the surrounding argument if it's available.
What it produces. A diagnosis of what the paragraph is doing in the argument, a choice of which critique layer to work at (foundation / structure / paragraph / sentence), and reasoned suggestions — never "change this," always "because X, therefore I'd suggest Y." Nothing is executed until you confirm, and every accepted change is logged.
Example exchange.
You: Take a look at this — "Stiegler's reading of Plato is, in a sense, deeply Heideggerian, and this matters for how we understand technics today."
Companion: Before I touch the wording, let me name what this paragraph is doing: it's an orienting claim — it's supposed to tell the reader why the Plato/Heidegger link matters. Right now it half-does that. Two flags, both above the sentence level:
- "in a sense" and "deeply" are hedges that blunt the claim. A reviewer reads them as you not being sure. What's the precise sense in which it's Heideggerian?
- "this matters for how we understand technics today" promises a payoff the paragraph never cashes. Either deliver it here or signal where you deliver it.
Want me to propose a sharper version once you tell me the precise sense? I don't want to guess your argument for you.
What it's for. Stepping back from the paragraph to read a whole chapter or full paper the way a careful reader would — top-down, argument first. This is the systematic, four-layer critique.
How to trigger it. "Read this chapter," "give me a full review," "review my section." You're handing over a large unit and asking for holistic judgment.
What it uses as input. The full chapter/paper, your style profile, your reader profile, and the four-layer critique model (Foundation → Structure → Paragraph → Sentence).
What it produces. A feedback report saved to _feedback/, with issues sorted into tiers — 🔴 blocker / major / minor / question. Crucially, it does not dump all of that on you at once. Following the attention-friendly rules, it gives you the total count and a category overview first, starts from the quick wins, and walks through 3–5 items per round. Confirmed fixes are batch-executed, and if the revision is large it snapshots a new version.
Example exchange.
You: I just finished Chapter 3 — give it a full review.
Companion: Read the whole chapter. Here's the shape before we dig in: 9 issues total — 2 blockers (both Layer 1, foundation), 3 structural, 3 paragraph-level, 1 sentence-level. I'm going to ignore the sentence stuff for now; no point polishing a paragraph whose argument might move.
Let's start with two quick structural wins, then hit the blockers:
- §3.2 and §3.4 make nearly the same point about periodization — one of them is redundant. (quick)
- The transition into §3.3 is missing; the reader falls off a cliff between the Foucault discussion and your archive material. (quick)
Confirm these two and I'll do them, then we'll sit with the first blocker — which is that the chapter's central claim isn't actually argued, only asserted, in §3.1.
What it's for. Going from an idea to prose — planning a new chapter, exploring argumentative directions, and then collaboratively drafting. Mode C is the front door to the four-stage "New Content Generation" flow (conception → development → draft → integration).
How to trigger it. "I want to add a chapter on X," "help me think through this new section," "let's start writing this," 继续写 / 帮我想想这个概念怎么展开.
Interaction posture — this is the heart of Mode C. Listening first, no rushing to a solution. The AI is midwife, not architect. Before any drafting, it presses you on the core ("if this chapter could keep only one sentence, which one?"), separates intuition from defensible claim, asks Socratic questions, and offers 2–3 possible argumentative paths rather than picking for you.
What it uses as input. Your spoken-rough idea, the conception/outline notes, your style profile, and the reference index.
What it produces. First a sharpened conception; then, in the drafting stage, an academic paragraph in your voice — but explicitly stamped [AI DRAFT — author to review], with sources marked and uncertainties flagged. The governing rule: what the AI writes is a "draft for discussion," not finished text. The more it drafts, the heavier your review burden, so it never quietly substitutes its thinking for yours.
Example exchange.
You: I want to write a new section arguing that media ecology misread McLuhan.
Companion: Good — but let me hold off on writing and listen first. One question to start: if this section could leave only one sentence in the reader's head, what is it? "Media ecology misread McLuhan" is a topic, not yet a claim — misread him how, and with what consequence?
And before we go further: is this a feeling you have from reading them, or a position you're ready to defend against a media-ecology scholar who'd say "we read him fine"? If it's the latter, who specifically is the misreader you have in mind?
Depending on your answer I see two or three paths — e.g., (a) a close re-reading of one McLuhan text against the standard ecology reading, or (b) tracing how the misreading propagated through Postman. I'd rather you pick than pick for you.
What it's for. Meeting the adversary before the journal does. This is, in the skill's own words, "the most valuable and the most courage-demanding function." It simulates real, hostile humanities review.
How to trigger it. "How would reviewers attack this?", "does this concept hold up?", "attack my argument." For the variant that targets your method rather than your claim, say "attack my method, not my claim."
What it uses as input. Your reader profile is read first — this is what makes the reviewers concrete rather than generic. It simulates three reviewers and one kind reader:
- Reviewer A · Theoretically demanding — tests whether your concepts are actually sharp.
- Reviewer B · Historically empirical — tests your evidence and watches for anachronism.
- Reviewer C · Methodologically skeptical — tests how you handle interdisciplinarity.
- Reader D · Well-intentioned but confused — not adversarial, just lost; the places D gets confused are often your real weak points.
Two things that make Mode D unusual.
- Calibration (1–5). You set the intensity, from 1 · gentle reader up to 5 · adversarial committee member (defense-rehearsal level). Default is 3 · peer reviewer. You can dial it mid-session ("drop to 2, I'm overwhelmed" / "go to 5") and it obeys immediately.
- Anti-sycophancy. LLMs cave when you push back; Mode D refuses to. It will concede a challenge only when at least 2 substantive conditions are met (you cited evidence, redefined scope, showed you'd already handled it, etc.). "I disagree" or "it's my style" does not count, and it'll say so.
What it produces. Concrete challenges in dialogue (not a verdict), and a record in the interaction log of every challenge → your response → status (addressed / still open) — so when the real reviewer raises it, your defense is already written.
Example exchange.
You: How would reviewers attack my claim that "the archive itself performs a kind of forgetting"? Push hard — level 4.
Companion: Level 4, adversarial. Reviewer A opens:
"'The archive performs forgetting' — performs in what sense? You're leaning on a personification that does real argumentative work for you. Either the archive is an agent (then defend that ontology) or 'performs' is a metaphor (then it can't carry the claim). Which is it?"
You: It's obviously metaphorical, that's just how this field talks.
Companion: That's not a rebuttal I can accept — "it's how the field talks" isn't a concrete basis, and leaning on convention is exactly what Reviewer A is probing. The challenge stands. If you mean something specific by the metaphor — a mechanism by which selection/cataloguing produces forgetting — name it, and the challenge stands down. Otherwise I'd address it head-on in the text rather than hope the reviewer reads it charitably.
What it's for. The most common state in academic writing isn't "help me revise this paragraph" — it's "I'm stuck, I can't write." Mode E exists for exactly that.
How to trigger it. "I'm stuck," "I don't know how to write this section," "I'm blocked," 我写不下去了. The skill also notices when you've been circling the same paragraph and may suggest it.
What it uses as input. Whatever you've got — the half-formed section, the conclusion you're aiming at, your reading materials. Often very little, which is the point.
What it produces. Movement, via five unblocking strategies:
- Dimensional reduction — replace the giant question with tiny ones ("what's the one thing you want the reader to take from this section?").
- Speak first — you talk, the AI takes notes, and together you turn speech into prose (true to "my hand writes my voice").
- Reverse engineering — start from the conclusion and work backward to the steps and materials needed.
-
Take another path — if the current line is blocked, open a fresh file in
_drafts/and try a different angle (a writing "feature branch" that doesn't touch main). - Reading supply — being stuck often means too little input; go find the missing source.
Example exchange.
You: I've been staring at the methods section for an hour. I can't write it.
Companion: Let's not write it yet — let's shrink it. Forget "the methods section." Answer just this: what's the one thing you most want the reader to trust about how you read these texts?
You: …I guess that I'm reading them in the original, not in translation, so the philological claims actually hold.
Companion: There's your opening sentence — and it's in your voice, not boilerplate. Say the next thing out loud the same way: why does reading the original change what you can claim here? I'll take it down as you go and we'll shape it into prose after, rather than you trying to write polished sentences cold.
What it's for. You have a draft — often one that's already been AI-polished — and you need to revise it systematically. The core challenge, in the skill's framing: keep the structural improvements of the draft while removing AI traces and restoring your own voice.
How to trigger it. "Help me revise this full draft," "clean up this AI-polished version," "I lost my voice in this." For the teaching variant, "teach me how to see this myself."
What it uses as input. Two versions — the draft (e.g., the AI-polished one) and an original (your early, un-AI'd manuscript, used as the true-voice anchor) — plus an established style profile and the ai-trace-checklist.md.
What it produces. A chapter-by-chapter revision where each change is judged: is this an improvement (clearer structure, better citation) or an alienation (lost voice, AI cliché)? Improvements stay; alienations get restored from the original or rewritten in your voice. Every edit is logged as restored original / kept draft improvement / rewritten. It deliberately works only 1–2 chapters per session — this is argument-level work, not a polish pass.
The F.coach sub-mode. Standard Mode F just hands you the revised text. F.coach refuses to — instead it asks you 3–5 diagnostic questions about the problem passage, waits for your answer, names what you noticed, and only then proposes a fix. It's slower on purpose (a 5-minute fix can take 20), because the goal is that you don't need the skill in five years. You can opt out anytime: "skip coaching, just give me the revision."
Example exchange.
You: Here's my chapter — the polished version and my original draft. Revise it but coach me through it.
Companion: F.coach it is. Looking at your opening paragraph: the polished version reads smoothly, but before I tell you what I think, answer these:
- In one sentence — what is this paragraph claiming?
- Your original opened with a question; the polished version opens with a thesis statement. Which one sounds like you?
- The polished version added the phrase "it is important to note that." Read it aloud — does that sound like anything you'd say?
You: Claiming that periodization is political. The question version is mine. And no — I'd never write "it is important to note."
Companion: Right — so you've already done the diagnosis. The polish smoothed a genuine move (opening on a question) into a flat thesis, and dropped in a filler phrase that isn't yours. My suggestion just confirms yours: restore the question-opening from your original, cut the filler. Logged as
restored original expression.
What it's for. A deliberately narrow, mechanical check: did the paper deliver on what it promised? What's unique here is that the AI turns its scholarly judgment off. It is not evaluating quality — only matching promises to deliveries. (Borrowed from Thesify's Purpose-Check design.)
How to trigger it. "Run blind reading," "did the paper deliver on its promises?", "I just finished Chapter 3, run blind reading," "one more promise-delivery pass before submission."
What it uses as input. Only the text's promises and deliveries. Notably, this mode deliberately does not read your _writing-config/ files — it steps outside your internal view on purpose, so it sees the paper the way a stranger would.
What it produces. A four-bucket report:
- ✅ Promises delivered — "this chapter will explore X" → and it does, here.
-
⚠️ Partially delivered — "answers in three parts A, B, C" → A and B appear, C never does. - ❌ Promises not delivered — "I'll return to scholar Z in the conclusion" → conclusion never mentions Z.
- 🤔 Implicit promises (AI inference, may be wrong) — a concept is introduced but never closed.
It will not evaluate whether a promise was worth making, will not rewrite undelivered promises, and will not write the missing sections. It just holds up the mirror — you decide whether to add the delivery or retract the promise.
Why it matters. Long writing causes promise drift: the intro's promises quietly get replaced by what you discovered while arguing, and you don't feel it happen. Blind reading puts the original promise and the actual delivery side by side.
Example exchange.
You: Finished the draft. Run blind reading before I submit.
Companion: Judgment off — I'm only matching promises to deliveries.
=== Blind Reading · Full draft === ## ✅ Promises delivered - "This paper will trace the concept across three thinkers" (intro ¶2) → Thinkers handled in §2, §3, §4. ## ⚠️ Partially delivered - "I respond to this in three parts: A, B, C" (§4 ¶1) → A in §4.1, B in §4.2 — no section addresses C. ## ❌ Promises not delivered - "I will return to Derrida's objection in the conclusion" (intro ¶5) → Conclusion does not mention Derrida. ## 🤔 Implicit promises (may be wrong) - §1 opens on the term "spectral," treated as central, but the term never returns. Intended?I'm not telling you to fix any of these — just whether to deliver C and Derrida, or to walk back those promises in the intro.
What it's for. The earliest-stage mode. You have an interest, a topic, a vague sense of what you want to say — but not yet a defensible research question. Mode H turns vague interest into a sharp, write-able question through Socratic dialogue.
Crucial: this is not PICO, not hypothesis-testing. Humanities questions have different shapes; generic AI scoping prompts produce empty, STEM-flavored questions. Mode H works strictly inside humanities conventions.
How to trigger it. "I want to do something on X" (vague), "help me sharpen my research question," 研究问题 — or before a proposal/prospectus, or when you're stuck between directions.
What it uses as input. Your interest, and _writing-config/discipline.md if it exists (otherwise it asks about your field).
What it produces. Across roughly 5–8 unhurried turns, it: locates the field, finds the puzzle (what's contested / undertheorized / over-saturated), names the question type (re-reading / re-construction / intervention), applies the "so what" test, identifies the real interlocutor (the strongest opponent — "the interlocutor is more important than the topic"), and sharpens the verb (from "explore" to "argue against / re-interpret / trace the genealogy of"). Output lands in _writing-config/research-question.md.
It will refuse to generate the question for you, and it will not approve a question until the "so what" test passes concretely. If you hand it a topic ("I want to write about Foucault and AI"), it declines to proceed until you narrow it to an actual claim.
Example exchange.
You: I want to write something about Arendt and social media.
Companion: That's a topic, not yet a question — and Mode H won't pretend otherwise. Let's sharpen it. First: what's contested here? Plenty has been written on Arendt and the public sphere online — where do scholars actually disagree?
You: I think people use her "public realm" too loosely for anything online.
Companion: Now we're moving. Try the so-what test — complete this: "If I show that the 'public realm' is misapplied to social media, then ___." And name your interlocutor: who is making the loose application you want to argue against? A question without a named opponent is still just a topic. Once you give me both, we'll pick the verb — and "explore" isn't on the menu.
What it's for. Organizing what you've already read into a working map of camps, lineages, and debates — and locating your position within it.
Iron rule: this mode does NOT search literature for you. AI lit-search breeds citation hallucination and replaces the irreducible work of reading. Mode I is downstream of reading, not a substitute for it.
How to trigger it. "Help me map my literature," "I've read these — how do they relate?", 文献综述 — typically before a lit-review section, or right after Mode H when the sharp question needs to be placed in a conversation.
What it uses as input. A list of references you have actually read — 8–15 minimum. If the list is shorter than 8, Mode I refuses to proceed and instead points you at more reading; it does not compensate for under-reading.
What it produces. Your references grouped by intellectual lineage / school / camp, a map of who agrees, opposes, or simply ignores whom, and an honest placement of your own position (sometimes you think you're in Camp A but your moves put you in Camp B — it'll surface that gently). It does gap detection by asking — "you cite X but not Y, who's usually read alongside X — deliberate?" — but it will never invent a scholar or summarize a work you didn't name. Output lands in _writing-config/literature-map.md. If one scholar is cited 3+ times and is load-bearing, it may suggest spinning up a dedicated perspective skill to use as a reviewer in Mode D.
Example exchange.
You: Map these for me: I've read Mumford, Ellul, Winner, Feenberg, and Stiegler.
Companion: Two flags before I map. First, that's 5 — under the floor of 8. Mode I won't build a real map from five names; it'd flatten a field you haven't read widely enough in yet. Want to keep reading first, or do a rough provisional sketch knowing it's incomplete?
Second, even provisionally: Mumford, Ellul, and Winner lean toward a substantivist/critical camp on technology; Feenberg explicitly tries to broker between substantivism and constructivism; Stiegler is doing something else again (pharmacological). Question for you — which camp are you writing from? And you haven't named anyone from the constructivist side (Bijker, Pinch). Deliberate intervention, or a gap? I won't tell you what they argue — but if you haven't read them, that's a reading task, not something I fill in.
What it's for. Pure outline mode — no draft writing. It was deliberately split out from Mode C so you can plan without accidentally tripping into drafting. It's discipline-aware: it pulls the standard rhetorical arc for your field.
How to trigger it. "Help me outline a chapter on X," "I need to plan this paper," "what structure should this take?", 写论文大纲 — and it's the natural next step after H + I (question sharpened, literature mapped).
What it uses as input. Your scope (chapter? full paper? dissertation chapter? book proposal? — length matters, a 6,000-word article and a 25,000-word chapter have totally different arcs), plus discipline.md, research-question.md, literature-map.md, and reader-profile.md where they exist.
What it produces. A section-by-section outline where each section gets (a) its function in the argument — "what does this section DO," not "what topic does it cover" — (b) key claim, (c) key evidence/source, (d) a word target. It applies a discipline-specific standard arc (e.g., Philosophy: concept-puzzle → analysis → defense against the strongest objection → downstream consequences; History: historiographical positioning → narrative → analytic argument → significance), then cross-checks every section against the research question — cutting sections that look like the topic but don't advance the question, adding steps the argument needs. Output lands in _writing-config/outline.md.
Hard constraint: Mode J does not write paragraphs. Ask it to "just write the first paragraph too" and it refuses, offering to switch to Mode C instead. The whole value of Mode J is the discipline of not writing.
Example exchange.
You: Outline my philosophy paper arguing that personal identity survives teleportation.
Companion: Scope first: journal article (~8,000 words) or dissertation chapter? The arc differs a lot. Assuming an article — philosophy's standard arc is concept-puzzle → conceptual analysis → defense against the strongest objection → consequences. So:
- §1 Function: open the puzzle — pose the teleportation case so the identity stakes are unavoidable. Claim: the case is a genuine dilemma, not a trick. ~900 words.
- §2 Function: analysis — distinguish numerical vs. qualitative identity. … ~1,500 words.
- §3 Function: defend against the strongest objection — the "no continuer" objection. This is where the paper lives or dies. …
One cross-check: your §4 as you described it is really about teleportation in fiction — interesting, but does it advance the identity claim? If not, I'd cut it. And no, I won't draft §1's first paragraph — that's Mode C. Want me to switch?
What it's for. Generating the AI-use disclosure statement that humanities journals increasingly require. Humanities policies are typically stricter than STEM ones — many ban AI co-authorship outright and require explicit tier disclosure — so a generic statement won't do.
How to trigger it. "Generate my AI-use disclosure," "what do I write about AI use?", before journal submission or dissertation deposit.
What it uses as input. Your paper trail: _meta/interaction-log.md, _meta/revision-log.md, any reflexive notes, and the skill's own record of which modes you actually used (especially C drafting and F revision). This is why the other modes log everything — so this statement is honest by construction.
What it produces. First, a tier classification on a humanities-specific 4-level scheme:
- Tier 0 · No AI (rare in 2026)
- Tier 1 · Proofreading / translation / formatting only — most journals accept without question.
- Tier 2 · AI as thinking partner / devil's advocate — no AI prose in the submission; most journals accept with disclosure.
- Tier 3 · AI-assisted prose — some prose was AI-drafted then revised; several journals restrict this, so the disclosure must be specific.
- Tier 4 · Substantial AI-generated prose — many top journals forbid this entirely, and Mode K warns you.
Then it checks your target journal's policy (and flags a mismatch), and generates a disclosure from template options A (short footnote) / B (standard methods paragraph) / C (detailed + per-section appendix), saved to _meta/AI-use-statement.md.
Hard constraints you can't talk it out of. It will not under-disclose — if you want to soften the statement it asks "what specifically, and why?", because the answer often reveals an ethical problem. It will not over-claim AI sophistication (journals read "the AI made critical contributions" as undisclosed co-authorship). And it always reminds you: you are responsible for everything in the submission, whatever the tier.
Example exchange.
You: Generate my AI-use disclosure. Keep it short — I don't want to draw attention.
Companion: I'll keep it as short as the truth allows, but I won't shrink it below your actual use. Reading your logs: you ran Mode D devil's-advocate on Chapter 3 (Tier 2), used Mode F to clean an earlier AI-polished version of Chapter 2 — and that chapter's §2.3 has three paragraphs logged as AI-drafted-then-revised. That last part is Tier 3, not Tier 2.
"Don't want to draw attention" worries me a little — so let me ask plainly: what specifically do you want to leave out? If it's the §2.3 drafting, note that several humanities journals restrict exactly that, and the honest move is to disclose it (or revise so it's no longer true), not to omit it. Here's a Template B paragraph that states it precisely without over-dramatizing — and remember, whatever we write, the arguments and final wording are yours to stand behind.
- 🏠 Home — the full wiki map
- New to the skill? The pre-writing trio H → I → J is the gentlest on-ramp.
- Already have a messy AI-polished draft? Jump straight to F.
- About to submit? Run G, then K.
humanities-writing-companion · CC BY-NC 4.0 · DOI 10.5281/zenodo.20280773 · by Shen Cong (沈聪)
📘 English
Guides
- Getting Started
- The 11 Working Modes
- Paper End-to-End
- Discipline Guides
- FAQ & Troubleshooting
- Citation & Writing Tools
About the project
📗 中文
指南
关于项目