chapter-humanizer

Chapter Humanizer

The chapter-humanizer is a surgical pass — not a rewrite. Its job is to find AI-construction patterns in already-reviewed prose and replace them with alternatives written in the author's voice. It does not touch craft, structure, story logic, or anything the chapter-reviewer already addressed.

Position in workflow: chapter-writer → chapter-reviewer → chapter-humanizer → chapter-proofreader → manuscript-checker

Prerequisites — MANDATORY LOADS

Before scanning a single line:

1. Draft — Read {project}/chapters/{chapter}/draft.md in full. If missing, stop and tell the user: "Kein draft.md für dieses Kapitel gefunden — chapter-writer muss zuerst laufen."
2. Anti-AI patterns — MCP get_craft_reference("anti-ai-patterns"). Why: Section 11 contains the shape catalog with banned-shape descriptions and examples. Section 1 contains the flagged-vocabulary list. Both are the scan targets.
3. Author profile — MCP get_author(). Why: Alternatives must be written in the author's documented voice (tone, rhythm, vocabulary). A proposed fix that doesn't match the author's profile is not a fix — it's a different kind of AI output.
4. Book CLAUDE.md — MCP get_book_claudemd(book_slug). Why: Book-level rules may contain additional banned shapes or construction constraints specific to this book.

Scan — Two Passes

Pass 1: Section 11 Elegant Abstraction Shapes

Scan every sentence for the following constructions. For each hit, record: line number (approximate), the offending text verbatim, the shape name.

Shape	What to search for
11.1 Word-count commentary	One word. / Two words. / Three words. followed by narrator editorial about the rarity of those words
11.2 Sentence-as-projectile	the words landed, the line landed, the sentence landed, settled into the room, sat in the room
11.3 Room-as-receiver	the room received, the silence held, the hall absorbed, the air carried, the quiet kept
11.4 Economic metaphor	most expensive [word/sentence/motion/gesture], the word cost him, paid in silence, [action] cost her (non-literal)
11.5 Near-miss body language	did not quite become, almost became a, never quite [verb] — flag count per scene
11.6 Body-part agency	[hand|breath|stomach|shoulders|face|mouth|eyes|chest|throat|jaw|spine|fingers|knee|feet|legs] + [had been|was|were|kept|started|began] + [deciding|having|choosing|wanting|refusing|trying|failing|knowing] — see Section 11.6 for full regex
11.6 Trust-split	trust his/her/my face, trust his/her voice, trust his/her hands + distrust/not trust variants
11.7 Backward-negation loop	what [pronoun] had been refusing/unable/failing to [verb] where the verb echoes the sentence's opening action
11.8 Expository repeat	Same noun-phrase or logical constraint appearing in two consecutive sentences — second sentence restates first to justify a narrative gap
11.9 Negation-as-assertion loop	It wasn't [X]. It was [Y]. or Not [X]. [Y]. — flag on 2nd+ occurrence per scene
11.10 Hedge-word density	seemed, appeared to, as if (non-simile use), might have — flag when 3+ combined instances per scene

11.5 scoring: One near-miss per scene is acceptable. Two or more in the same scene → flag all instances.

11.9 / 11.10 density rules: Count instances per scene. 11.9: flag from the second negation-assertion per scene onward. 11.10: flag when seemed / appeared to / as if (non-simile) / might have reaches 3+ combined per scene — report the count and all instances.

Pass 2: Flagged Vocabulary (Section 1)

Scan for the 60 flagged words and phrases from Section 1 of anti-ai-patterns.md. Entries 1–55 include core AI-vocabulary (delve, tapestry, nuanced, vibrant, landscape (metaphorical), embark, resonate, pivotal, realm, testament, intricate, myriad, unprecedented, foster, navigate (metaphorical), etc.). Entries 56–60 are formal transition tells: Furthermore, Moreover, In addition, Conversely, On the other hand.

For each hit: record word, sentence, context. Do not flag words that are clearly literal or in-character dialect.

Also check author profile's writing_discoveries.donts and vocabulary.md banned list — these are book/author-specific additions.

Output: Scan Report

After both passes, present the findings as a numbered list. Do NOT apply any changes yet.

## Humanizer Scan — {book-slug} / {chapter-slug}

**Section 11 shapes found: N**
**Flagged vocabulary hits: M**
**Near-miss count per scene: [Sc1: X, Sc2: Y, ...]**

---

### Hits

[1] **11.6 Body-part agency** — Sc 2, ~line 47
> *"His throat had been refusing to close on what he'd been holding since yesterday."*
Proposed fix: *"His throat tightened. Yesterday he hadn't been able to cry. He still couldn't, not entirely."*

[2] **11.6 Trust-split** — Sc 2, ~line 52
> *"stayed there until he could trust his face again."*
Proposed fix: *"He held still. When he was sure his expression had settled, he moved."*

[3] **11.8 Expository repeat** — Sc 3, ~line 89
> *"They had silver on her and weren't careful about the rest. They did not look under the shirt of a vampire they had silver on."*
Proposed fix: *"Silver was enough. They didn't check further."*

[4] **Flagged vocabulary** — Sc 1, ~line 12
> *"...the nuanced tension between them..."*
Proposed fix: *"...the tension between them, which neither named..."*

---

**Instructions:** Reply with the hit numbers you want to apply as-is, numbers you want a different alternative for (e.g. "3: shorter"), and numbers you want to skip. Example: "apply 1, 2, 4 / rework 3: make it one sentence / skip none"

Do not present more than 20 hits at once. If the chapter has more, present the first 20, apply those, then continue with the next batch.

Interaction Loop

User Response Formats

The user responds with one of:

• apply all — Apply every proposed fix as-is.
• apply N, M, ... — Apply specific hit numbers as-is.
• skip N, M, ... — Leave those hits unchanged.
• N: [instruction] — Rework hit N with the given instruction before applying.
• apply N, M / skip P / Q: shorter — Mixed response.

Rework

When the user requests a rework (N: [instruction]), generate a revised alternative that:

• Addresses the specific instruction
• Stays in the author's documented voice
• Does not introduce new Section 11 shapes or flagged vocabulary

Present the rework for confirmation before applying: "Rework for [N]: '[revised text]' — ok?"

Applying Changes

After the user approves (or applies with reworks confirmed):

1. Read the full draft.md again before writing (GH#27 — file may have changed if the session has been long).
2. Apply ALL approved changes in a single write pass — do not write the file multiple times.
3. Report: "Applied N changes. Skipped M. Draft updated."
4. If flagged vocabulary was accepted to stay, note it: "[word] kept at your request."

Iteration

After applying, offer: "Möchtest du noch eine Runde? Oder weiter zu /storyforge:chapter-proofreader?"

If the user wants another pass: re-scan the updated draft (the fixes may have introduced new issues — rare but possible). Cap at 2 iterations per session.

Rules

• Surgical only. Change only the flagged construction. Do not improve surrounding prose, fix style, or add content. The chapter-reviewer already handled craft.
• Author voice is mandatory. Every proposed alternative must match the author's documented tone, rhythm, and vocabulary. An alternative that sounds like a different author is a regression, not a fix.
• No wholesale rewrites. If a passage has so many Section 11 shapes that individual fixes would require reconstructing the scene, stop and tell the user: "Szene [N] hat [X] overlapping shapes — eine gezielte Überarbeitung der ganzen Szene wäre effizienter als Einzelfixes. Soll ich Vorschläge für die ganze Szene machen?" Then wait for explicit confirmation before proceeding.
• Do not create new AI-tells. Before proposing any fix, run a quick mental check: does the proposed alternative introduce a new Section 11 shape, flagged vocabulary word, or other known tell?
• Read the full file before writing. GH#27 applies here. Always re-read draft.md before the write pass.
• voice-checker is optional after humanizing. If the user wants a holistic score after this pass, suggest /storyforge:voice-checker. But the humanizer's targeted pass is more actionable for the patterns it covers.