Narrative

Narrative v2 — Craft + Slop Defense + Project Pipeline

Assist with narrative prose: fiction, lore, character backstories, quest text, dialogue, worldbuilding — and full short-novel projects. Grounded, buildable writing. Not abstract exposition.

Reference files (verbatim, do not edit):

• references/autonovel-craft.md — plot / character / worldbuilding / foreshadowing, each with a mechanical check.
• references/autonovel-anti-slop.md — banned words, structural slop, tone, detection signals.
• references/project-mode.md — manuscript tree, schemas, gates (this skill's project extension).

Load with skill_view(name='narrative', file_path='references/…') when you need exact rule text.

Modes

In project mode, refuse chapter assembly and export if ledger/canon/voice are missing or scenes are not reviewed.

Skill tests / short samples stay scene-scale. Do not expand into a full book pipeline unless the user asks.

Single-scene procedure (v1, retained)

1. Anchor tone — name register (dread / restraint / melancholy / bright action). Ask if vague; use tone defaults above for this user's fantasy.
2. Structure — one framework from CRAFT §1 (Save-the-Cat | Harmon circle | Sanderson PPP). Prefer MICE + try-fail: middle beats mostly "yes, but…" / "no, and…".
3. Character before prose — Three Sliders (proactivity / likability / competence — ≥2 high, or one high with growth) + Wound→Lie→Want→Need (Want ⊥ Need).
4. System / magic — Three Laws (CRAFT §3): hard rules, limitations ≥ powers, costs drive decisions, no unforeshadowed power in final 25%.
5. Draft show-don't-tell (CRAFT §5) — concrete nouns, earned metaphor, rhythm variation (short → long → short). Zero telling at peaks.
6. Dual-persona review — A literary critic (prose, voice, monotony); B fiction professor (character, ethics, beats). Fix top items until "mostly qualified hedges."
7. Kill slop (ANTI-SLOP) — Tier-1 zero; Tier-3 delete; em dashes ≤2/page; kill "not just X, but Y"; break topic-sentence templates.
8. Stability-trap check (mandatory) — all 7 points below.

Project mode

Init

1. Scaffold manuscript/{project}/ from templates/ (or book-pipeline templates — same fields).
2. Write concept.md with user (logline, tone, MICE, target length 40–50k default).
3. Seed character sheets, plot-ledger.md, foreshadow-bank.md, worldbuilding.md, canon.md.
4. One-time voice calibration → voice-profile.md (see book-pipeline style-revision playbook §3).
5. Bind via optional manuscript.yaml (title, framework, target_words, voice_profile).

Continuity ledger (read every draft/review)

Generator loads only what it needs for the current scene (beat row + relevant characters + last chapter summary) — never the full manuscript.

Draft (generator stage)

1. Read assigned beat from ledger + voice-profile + relevant character rows.
2. Run single-scene procedure steps 1–5.
3. Frontmatter on scene/chapter file:

   ---
   scene_id: ch03-sc02
   beat: catalyst
   try_fail: no-and   # yes-but | no-and | no-but | yes-and
   pov: Jin
   status: drafted
   ---
   

4. Do not self-critique for stability trap at book scale — hand to reviewer.

Review (separate stage — use narrative-revisor or style-revision playbook)

Fresh context. Loads ledger + canon + voice + one scene. Never generator reasoning.

• Pass 1 craft/structure
• Pass 2 anti-slop mechanical
• Pass 3 stability-trap 7-point
• Emit JSON handoff (scene_id, issues[], severity)

Revise

Drafting agent applies only reviewer issues. Re-submit until pass status with zero blocker/major.

Assemble & chapter gates

Before marking a chapter done:

1. Fact continuity — new facts added to canon.md
2. Ledger progression — beats moved planned→drafted→reviewed; no silent skips
3. Foreshadow — no dangling past payoff window; no unforeshadowed climax element
4. Character state — sliders/arc match sheet
5. Stability-trap at chapter grain
6. Anti-slop per chapter (Tier-1 zero, em-dash cap)

Export (via book-writer-pipeline CLI) only when all chapters done.

Scriptable checks

Run scripts/check_manuscript.py <manuscript_dir> for mechanical scans (Tier-1, em-dash density, dangling plants, beat status). Judgment items (moral ambiguity, emotional range) stay human/reviewer.

Stability Trap — non-negotiable (all scales)

If a draft rounds a sharp edge safer/generic → rewrite.

Anti-slop (project + single)

• Tier-1 kill on sight; Tier-2 ≤2/paragraph; Tier-3 delete all filler
• Em dashes ≤2/page; zero "not just X, but Y"
• No topic-sentence machine, list abuse, hedge parade, transition addiction
• Fiction AI tells (CRAFT §6): "a sense of…", "eyes widened", "wave of emotion", etc. — kill

Full lists: references/autonovel-anti-slop.md.

Pitfalls

• Lists instead of prose for weighty scenes
• Throat-clearing openings ("In this land…")
• Info-dump >100 words without action/dialogue
• Self-review at book scale (stability trap doubles)
• Ledger drift / voice drift across chapters
• Dumping full manuscript into a 16K–32K context
• Expanding a skill test into a pipeline without being asked

Verify

• Read aloud — person who has lived, not press release
• Tier-1 search = zero
• ≥1 surprising sentence
• Stability 7/7 at the grain you shipped (scene / chapter / book)
• Project mode: ledger statuses match reality; no dangling plants at export

Book-scale ownership

This skill owns craft inside a scene/chapter. The book-pipeline skill owns architecture, folder convention, export, publish routing, and multi-lane orchestration. The book-writer profile owns the loop end-to-end.

When book work spans craft + models + CLI + market + publish: delegate lanes. Do not solo the whole novel from one profile.