-
Notifications
You must be signed in to change notification settings - Fork 0
End to End Workflow
Most AI writing tools meet you in the middle of a sentence. The Humanities Writing Companion meets you at the beginning of an idea — and stays with you until the moment you press submit. This page is a walkthrough of that whole arc: one paper, taken from a vague hunch to a finished manuscript, using the modes in the order a real project tends to need them.
The story is deliberately concrete. We'll follow a graduate student — call her Lin — writing a journal article in the history/philosophy of technology. Her starting point is nothing more than: "I find Stiegler's idea of the pharmakon interesting, and I think it says something about AI." That is not a paper. By the end of this page it will be.
For the full reference on each mode, see The 11 Working Modes. Here we care about the journey — what you do at each stage, which mode picks it up, and what comes back.
A note before we start: the modes are not a rigid pipeline. The skill escalates and de-escalates between them on the fly — if a planning conversation reveals a hole in the argument, it'll pull you sideways into the devil's advocate; if a draft needs rewriting rather than tweaking, it switches from revision back to conception. The sequence below is the spine. Real projects loop.
Before any of the lettered modes run, the first conversation sets up the workspace. Lin says "I want to write a paper on Stiegler and AI," and the skill asks the questions that determine everything downstream:
-
What are you writing, and in what discipline? This is the load-bearing question. Discipline isn't metadata here — it's a routing variable. Lin's work is intellectual history doing philosophical argument, so it inherits both History and Philosophy critique dimensions. That declaration gets written to
_writing-config/discipline.mdand consulted on every single critique afterward. - Citation format? Chicago, for a humanities journal.
- Target reader? A reviewer in philosophy of technology who is sympathetic to Stiegler but allergic to loose conceptual borrowing. This goes into a reader profile — and later it becomes a named adversary.
- Any existing writing? Lin pastes two paragraphs from an old seminar paper. The skill reads them to build a style profile: her habit of comma-flowing long sentences, her first-person voice, her preference for keeping certain Greek terms untranslated.
These three files — discipline, reader profile, style profile — are the spine of memory. They persist across sessions, so when Lin comes back three weeks later and says "let's keep working on the Stiegler paper," nothing has to be re-explained.
Now the lettered journey begins.
What Lin does: arrives with an interest, not a question. What she gets back: a single sharp sentence she's willing to defend.
Lin's opening line — "the pharmakon says something about AI" — is a topic, not a research question. Mode H exists precisely to refuse that and push for more. Crucially, it does not do this the way a generic AI scoping tool would. There's no PICO, no hypothesis, no STEM-flavored "variables." Humanities questions have their own shapes, and Mode H knows them.
Over six or seven turns of Socratic back-and-forth, the skill works through a checklist with Lin rather than for her:
- Where's the puzzle? Not "what's interesting" but "what's contested, undertheorized, or over-saturated?" Lin realizes the over-saturated move is "AI is a double-edged sword" — everyone says that. The undertheorized move is what Stiegler actually meant by pharmakon, which is sharper and stranger than the cliché.
- What type of question is this? The skill names three humanities shapes — re-reading a classic against the grain, re-constructing a tradition, or intervening in a current debate. Lin's is an intervention: bringing Stiegler's conceptual apparatus into the live debate about AI.
- The "so what" test. Lin has to complete the sentence: "If I answer this, then ___." Her first attempt is mushy. The skill won't approve the question until she can fill the blank concretely — and eventually she does: then the "use AI carefully" framing that dominates the ethics literature is revealed as resting on a notion of the user that the pharmakon dissolves.
- Who's the real interlocutor? This is the move the skill insists matters most. A question without an interlocutor is a topic, not a research question. Lin's interlocutor turns out to be exactly the reader from onboarding: the technology-ethicist who thinks careful use is the answer.
- Sharpen the verb. "Explore" and "examine" are banned. Lin commits to "I will argue that…".
What lands in _writing-config/research-question.md is no longer a vibe. It's something like: "I argue that the dominant 'use AI responsibly' framing presupposes a user who stands outside technology — a position Stiegler's pharmakon structurally forecloses." That sentence will discipline everything that follows.
What Lin does: lists the 12 works she's actually read. What she gets back: a map of the camps, and an honest read on where she actually stands.
Now the question needs to be located in a conversation. Mode I has one iron rule, stated up front and repeatedly: it does not search the literature for you. Asking an AI to find sources is how you get hallucinated citations and skip the irreducible work of reading. Mode I is strictly downstream of reading — it organizes what's already in Lin's head.
Lin lists her dozen sources. (Had she listed fewer than eight, the mode would have refused to proceed and sent her back to read more — it will not paper over under-reading.) The skill then does the thing that's genuinely hard: it groups the scholars by intellectual lineage, not by topic. Who would cite whom approvingly? Who opposes whom? It sketches the camps — the Stiegler-aligned organology readers, the mainstream AI-ethics "responsible use" camp, the posthumanist agency theorists who are technically in a different language game.
Two moments make this stage worth it:
- It locates Lin honestly. She thinks she's writing from the Stiegler camp. But the skill gently flags that several of her actual moves — her appeals to user responsibility — put her closer to the camp she means to critique. Surfacing that early saves a reviewer from doing it later, less kindly.
- Gap detection without invention. The skill notices she cites one organology scholar but not the obvious companion usually read alongside — and asks whether that's deliberate. It never asserts what the missing work says (that would be hallucination); it only asks if she considered it. If she hasn't read it, she goes and reads it. The AI does not fill in.
There's also a nice optional hook here: because Lin cites Stiegler five times and his framework is load-bearing, the skill suggests generating a dedicated stiegler-perspective lens she can later deploy as a reviewer in the devil's-advocate stage.
The output, _writing-config/literature-map.md, records the camps, her position, the specific debates her paper joins, and an honest list of "things I should read more of."
What Lin does: asks for a structure. What she gets back: a section-by-section outline where every section has a job, not just a topic — and not a single drafted paragraph.
This is where the skill's discipline shows. Mode J is plan-only, and it means it. If Lin says "while you're at it, just write the first paragraph too," it refuses and offers to switch to drafting mode instead. The entire value of Mode J is the discipline of not writing — of staring at the skeleton before putting flesh on it.
It reads her discipline file and applies a discipline-specific arc. Because she's doing intellectual history, the template is roughly: method declaration → context reconstruction → text analysis → concept-migration narrative. Lin can deviate, but now any deviation is a choice, not an oversight.
Then it builds the outline section by section, and for each section it forces four things: (a) its function in the argument — "opens the puzzle," "establishes the gap," not "discusses Stiegler"; (b) the key claim; (c) the key evidence or text; (d) a rough word target. Function is the discipline here: what does this section DO for the argument, never what topic does it cover.
Finally it cross-checks the outline against the research question from Stage 1: does every section serve the question? Is there a section that looks like the topic but doesn't advance the argument? (Cut it.) Are there steps the argument needs that aren't in the outline yet? (Add them.) The result lands in _writing-config/outline.md with an "argument trace" at the bottom — claim 1 supports claim 2 supports the thesis — so the logical spine is visible at a glance.
What Lin does: talks through what each paragraph is really trying to say. What she gets back: draft paragraphs in her voice, explicitly marked as drafts for her to own.
Now, finally, prose. Mode C is the entry point to actual writing, and its defining posture is stated bluntly: the AI is a midwife, not an architect. It listens first.
Before any drafting, there's a clarification step unique to Mode C. The skill presses: "If this section could leave only one sentence, which sentence?" It distinguishes a feeling from a defensible claim. It offers two or three possible argumentative paths rather than picking one for her.
Then comes the collaborative drafting flow, and this is the heart of "my hand writes my voice":
- Lin speaks first. Even roughly — "this paragraph, I want to say Stiegler took the pharmakon from Derrida's reading of the Phaedrus, and the point is that the cure and the poison are the same thing at the same time, not a choice."
-
The skill expands that into an academic paragraph — using her style profile (the comma-flow, the first person, the untranslated Greek), marking a source for every citation, and stamping the whole thing
[AI DRAFT — author to review]. Where it's unsure about a concept or a citation, it flags the spot with>>>rather than papering over the uncertainty. -
Lin revises. She rewrites, and the skill quietly records the pattern of her edits as a
[author micro-adjustment]— learning, over time, to sound more like her. - Style verification closes the loop.
The principle the skill keeps repeating: the more text it drafts, the heavier Lin's review burden. It must never quietly replace her thinking. A drafted paragraph is a proposal, not a deliverable.
By now Lin has a real first draft of a section — built on her own articulation, in her own cadence, with every borrowed idea attributed.
What Lin does: says "read this whole section." What she gets back: a top-down, four-layer diagnosis sorted into a triage report.
A draft exists; now it gets read properly. Mode B runs the skill's signature four-layer critique, strictly top-down:
Layer 1 · Foundation Does this 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 cardinal rule: do not exert effort at a lower layer while a higher layer is unresolved. If a paragraph's premise is broken (Layer 1), the skill will not polish its commas (Layer 4) — it diagnoses the foundation first, waits for Lin to decide direction, and only then descends.
Here's that rule biting in practice. Suppose one of Lin's paragraphs still reads:
Stiegler's pharmakon shows technology is both poison and remedy. AI is therefore a pharmakon too. We must be careful in our use of AI. As Stiegler said, technology is a double-edged sword.
The skill's response stops at Layer 1 and refuses to go lower:
🔴 Layer 1 · Foundation — Argument standing
This paragraph equates pharmakon with "double-edged sword." Stiegler
explicitly rejects that: the pharmakon is curative AND destructive at the
SAME TIME — and the destructive force is proportional to the curative one.
"Double-edged sword" smuggles in a user-choice framing; pharmakon names a
structural condition that precedes user choice. That difference is the
whole reason the concept is worth borrowing.
Verdict: Foundation broken. Do not proceed to lower layers until fixed.
🟡 Layer 2 · Structure — (deferred until L1 fixed)
🟢 Layer 3 · Paragraph — (deferred)
⚪ Layer 4 · Sentence — (deferred)
The feedback report saves to _feedback/ and sorts everything into four tiers borrowed from code review — 🔴 Blocker, 🟡 Major, 🟢 Minor, and the crucial ❓ To discuss (things that aren't the AI's to decide, like whether to widen the scope of the central claim). And because the skill is attention-aware, it doesn't dump twenty items at once: it gives the total count and category overview, starts from quick wins, and works through 3–5 items per round.
What Lin does: asks "how would a reviewer attack this?" What she gets back: a real adversary — calibrated to how much she can take today — that won't fold the moment she pushes back.
This is the mode the skill calls its most valuable and most courage-demanding. Mode D simulates a defense committee — three reviewers and one well-meaning-but-confused reader:
- Reviewer A · Theoretically demanding — tests the sharpness of concepts. Why is your reading of pharmakon different from the standard one? Is your synthesis a real synthesis or a salad?
- Reviewer B · Historically empirical — tests the historical narrative. Did the actors themselves use these terms? Primary or secondary sources?
- Reviewer C · Methodologically skeptical — tests coherence across the history/philosophy straddle. How deep is your competence in each field you're crossing?
- Reader D · Well-intentioned but confused — not adversarial, just lost. The distinctive design here: the places where a friendly reader trips are often the real weak points — more revealing than any attack.
Two features make this more than theater:
Calibration (1–5). Before launching, the skill asks how hard Lin wants to be hit today. Level 1 is a gentle reader for a fragile early draft; Level 3 (the default) is a standard peer reviewer; Level 5 is an adversarial committee member who will press to fail — for defense rehearsal, so she fails in private rather than in public. She can dial it down mid-session ("ease off, I'm overwhelmed") and the skill complies instantly, no negotiation. Because the reader profile from onboarding is loaded, Reviewer A isn't generic — it is the specific technology-ethicist Lin named in Stage 0.
Anti-sycophancy. LLMs tend to cave the moment you push back. Mode D is built against exactly that. When Lin pushes back on a challenge, the skill concedes only if at least 2 of 5 substantive conditions are met — she cited specific evidence, redefined the concept's scope, showed she'd already handled it in a footnote, and so on. "I just disagree" or "that's my style" does not count. (At Level 5 the bar tightens to 3 of 5.) Every resolved challenge is logged — "challenge → response → addressed" — so when a real reviewer raises the same point, Lin already has her answer drafted.
There's also a methodology-focus sub-mode that attacks only her moves, not her claims — the kind of attack that surfaces root causes rather than symptoms. For Lin's intellectual-history work it would press: Begriffsgeschichte or Cambridge School? Are you judging Stiegler by present concerns? Have you paid the cost of migrating the concept across periods?
What Lin does: admits "I'm stuck — I've been circling this section for an hour." What she gets back: not a paragraph handed to her, but a way back into motion.
The most common state in academic writing isn't "fix my paragraph" — it's "I can't write." Mode E treats the block itself as the problem and offers five unblocking strategies, picking whichever fits:
- Dimensional reduction — drop the impossible big question ("what should this section contain?") for a small one: "What's the one thing you most want the reader to take from it?"
- Speak first — Lin talks it out, the skill takes notes, and together they turn speech into prose. (This is the same "my hand writes my voice" move from Mode C, repurposed as an unsticking tool.)
- Reverse engineering — start from the conclusion and work backward to the steps that reach it.
-
Take another path — if the current line is genuinely blocked, open a fresh file in
_drafts/and try a different angle. It's the writing equivalent of a feature branch: experiment freely without touching the main draft. - Reading supply — sometimes a block just means not enough input yet, and the fix is to go back to the sources.
Lin's stuck section unsticks via reverse engineering: once she states the conclusion the section must reach, the three intermediate steps become obvious, and she's writing again.
What Lin does: brings an AI-polished version of a chapter she doesn't fully trust, plus her original. What she gets back: a revision that keeps the genuine improvements and surgically removes the parts that stopped sounding like her.
Somewhere along the way Lin ran an earlier chapter through a generic polishing tool, and it now reads smoother but somehow not like her. Mode F is the dedicated workflow for exactly this: it works from two versions at once — the polished draft and her original early manuscript — and its whole challenge is to keep the structural improvements while stripping the AI traces and restoring her voice.
Chapter by chapter, it compares the two and judges every single change: is this an improvement (clearer structure, better citation form) or an alienation (lost voice, an imported cliché)? Improvements stay. Alienations get reverted to her original phrasing or rewritten in her cadence. It runs the AI-trace checklist item by item — hunting the tells: "It is worth noting," functionless "Furthermore"s piled up, over-tidy parallelism that humanities prose never actually grows naturally, her first-person quietly swapped for an objectivized "this paper." Then the read-aloud test: does the revised paragraph sound like Lin?
There's also a teaching variant, Mode F.coach, for when Lin says "teach me to catch this myself." Instead of handing her the fix, the skill issues 3–5 diagnostic questions at the right layer ("read it aloud — where does your breath pause? does the punctuation match?"), waits for her answers, and only then proposes a revision — by which point she's usually seen it herself. It's slower on purpose. The trade is that in five years she shouldn't need the skill at all.
What Lin does: says "blind-read this before I call it done." What she gets back: a cold, mechanical ledger of every promise the paper made versus what it actually delivered.
This mode is unlike all the others, and deliberately so: the skill turns its scholarly judgment off. It does not evaluate quality. It does not read the style profile or any of the _writing-config/ files — it steps outside Lin's internal view on purpose. It does exactly one thing: checks whether the paper did what it said it would do.
Long writing causes promise drift — the questions promised in the introduction quietly get replaced by what the author discovered while arguing, and the author can't feel it happening. Mode G is the mirror. It mechanically extracts every promise ("this paper will…," "I will return to X's critique in the conclusion," "this section proceeds in three parts: A, B, C") and then checks each one for delivery, with no judgment of how well:
=== Blind Reading · Stiegler/AI paper ===
✅ Promises delivered
- "I will argue the responsible-use framing presupposes an outside user" (intro ¶3)
→ delivered: §4 ¶2–4 dissolve the user position directly
⚠️ Partially delivered
- "this section answers in three parts: A, B, C" (§3 ¶1)
→ A in §3.1, B in §3.2 — no section for C
❌ Not delivered
- "I will return to the technology-ethicist's objection in the conclusion" (intro ¶5)
→ not found — conclusion never circles back
🤔 Implicit promise (AI inference, may be wrong)
- §2 introduces "organology" but never returns to it — should there be closure?
Crucially, it does not write the missing pieces or tell Lin which promise to keep. It only holds up the gap. Whether to add the delivery or retract the promise is hers to decide. Lin spots immediately that she dropped the conclusion's return to her interlocutor — the exact thing a reviewer would have pounced on.
What Lin does: asks "what do I write about AI use for this journal?" What she gets back: an honest, tier-accurate disclosure statement — and a warning if her intended journal won't accept her actual tier.
The paper is done. One step remains that most generic tools ignore entirely: humanities journals increasingly require an AI-use disclosure, and their policies are typically stricter than STEM journals — many forbid AI co-authorship outright and require explicit tier disclosure.
Mode K audits the actual record — the interaction log, the revision log, the modes Lin really used — and classifies her AI involvement on a humanities-specific four-tier scheme:
- Tier 1 — proofreading / translation / formatting only. Accepted almost everywhere.
- Tier 2 — AI as thinking partner and devil's advocate, but no AI-drafted prose in the submission. Accepted by most humanities journals with disclosure.
- Tier 3 — some prose was AI-drafted then author-revised. Several journals restrict this; disclosure must be specific.
- Tier 4 — substantial AI-generated prose. Many top journals forbid this entirely — the skill warns her.
Lin used Mode C drafting (then heavily revised) and Mode D dialogue, which puts her at Tier 2–3. The skill generates a disclosure paragraph naming each specific use — Socratic work on the research question, devil's-advocate stress-testing in §3, AI-trace cleanup of the polished chapter — and, where prose was AI-drafted, says exactly where and how much.
Its hard constraints are pointed. Do not under-disclose: if Lin wants to soften the statement, the skill asks what she wants to remove and why — and notes that the answer often reveals an ethical problem rather than a wording one. Do not over-claim AI sophistication either, because a journal reads "the AI made critical contributions" as a confession of co-authorship. And it reminds her of the bottom line every time: the author is responsible for everything in the submission, regardless of tier. The statement saves to _meta/AI-use-statement.md, and Lin submits.
That's the spine, start to finish. But notice how often the real path looped: the Stage 5 review escalated into a Stage 6 stress-test; the Stage 6 methodology attack sent Lin briefly back to sharpen a claim; the bottleneck in Stage 7 reused the drafting move from Stage 4. The skill is built to move between modes mid-conversation — escalating from a paragraph to a chapter when a local problem turns out to be structural, de-escalating from review back to revision once direction is set — and to tell you when it switches, so you always know which altitude it's working at.
The throughline across all ten stages is the same opinionated stance: the AI reads for you, it doesn't write for you. It sharpens the question without answering it, maps the literature without inventing it, drafts only what you've articulated, attacks without caving, and discloses without flattering. Prose is the argument — so the voice stays yours.
| Stage | You (the author) do… | Mode | You get back… |
|---|---|---|---|
| 0 · Onboarding | Declare discipline, citation style, reader, paste a writing sample | — | Persistent discipline / reader / style profiles |
| 1 · Sharpen the question | Arrive with a vague interest | H | One sharp, defensible question + named interlocutor |
| 2 · Map the literature | List 8–15+ works you've actually read | I | A map of the camps + an honest read on your position |
| 3 · Plan the paper | Ask for a structure (no writing) | J | Section-by-section outline; every section has a job |
| 4 · Conceive & draft | Say what each paragraph means, in your words | C | Draft paragraphs in your voice, marked [AI DRAFT]
|
| 5 · Chapter review | "Read this whole section" | B | Top-down 4-layer triage report (Blocker/Major/Minor/❓) |
| 6 · Stress-test | "How would a reviewer attack this?" | D | A calibrated (1–5) adversary that won't cave |
| 7 · Unblock | "I'm stuck" | E | Five strategies to get moving again |
| 8 · Revise draft | Bring polished + original versions | F | Improvements kept, AI traces stripped, voice restored |
| 9 · Blind read | "Did it deliver on its promises?" | G | Mechanical promise-vs-delivery ledger |
| 10 · Disclose | "What do I write about AI use?" | K | Tier-accurate disclosure statement |
For the detailed specification of each mode — calibration levels, the four critique layers, the anti-sycophancy conditions, and the discipline-routing rules — see The 11 Working Modes.
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