---
title: "Manga Monthly Chapter — 60-panel editorial-tier chapter with a 5-LoRA cast"
reliability: 3.0 # LoRA .8 + panels img .3 + cover img .3, 7 happy-path stages (−.4), page-assembly is skill-side not a registered cap (−.2 judgment); with OPTIONAL teaser (i2v .6 + music .2) ≈2.2; no shipped showcase
---

# Manga monthly chapter — agent playbook ⭐ flagship

> Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. **This is the multi-LoRA flagship for manga / Webtoon serial IP creators.** A 60-panel monthly chapter — black-and-white halftone screentones, multi-character ensemble on-modelness via 5 trained LoRAs, table-of-contents, cover, and a chapter PDF ready for Shogakukan / Shueisha editorial submission or direct publication on Tapas / Webtoon. The Shōnen-Jump-pre-submission tier of work.

## What you'll get

A monthly-chapter-grade deliverable, structured the way a manga editorial team expects a chapter to land:

- **5 trained character LoRAs** — protagonist, antagonist, 2 supporting cast, 1 key location. Trained once per series, reusable forever. The Shōnen-Jump on-modelness bar by training, not by reference.
- **A series brand kit** — halftone-screentone B&W style, panel-border style, dialogue font, font for SFX, finishing chain (page-assembly, gutters, ToC, cover, PDF export).
- **60 panels** — multi-character ensemble shots routed automatically (protagonist holds, antagonist holds, supporting holds, location holds — all in the same panel). Drift gate at 0.88 (tighter than the pilot's 0.85 — editorial-submission tier).
- **Auto-page-assembled chapter** — the 60 panels arranged into pages with gutters, dialogue placed, SFX placed, panel-border style enforced. Reading-direction-aware (RTL for traditional manga, LTR for Webtoon).
- **Generated cover** — your most kinetic panel re-rendered at editorial-cover spec (B5 hardcopy + 1080×1620 digital thumbnail).
- **Table of contents** — chapter title, panel-count, page-count, generated from the chapter's own metadata.
- **Editorial-ready chapter PDF** — assembled pages + cover + ToC, ready to submit to a manga editorial team.
- **Webtoon single-scroll** — same panels stitched vertically for Tapas / Webtoon Canvas direct upload.
- **9:16 1-min social teaser** — strongest 8 panels with subtle motion + brand audio.

**Time**: 1 day for chapter 1 (heavy training + brand kit setup); 4-6 hours for follow-on chapters. **Cost**: $30-50 per chapter + one-time ~$20 for the 5-LoRA cast (Week 0). *This is the editorial-tier flagship* — the difference between "Webtoon Canvas indie" and "Shogakukan pre-submission."

## Tell the agent about the chapter

```yaml
series_title:        # e.g. "Tetsuken no Saga". Used in chapter card + brand kit name.
series_logline:      # 1-sentence pitch. Sets the editorial tone.
chapter_number:      # e.g. 3. Used in chapter card + ToC + filenames.
chapter_title:       # e.g. "Chapter 3 — The Iron-Spire Duel".
chapter_brief:       # 1-paragraph synopsis of THIS chapter's arc. Drives the 60-beat plan.
protagonist_name:    # Lead. Becomes LoRA 1 + appears in the cast roster.
protagonist_refs:    # OPTIONAL on chapter 2+. On chapter 1: 20-25 reference images of the lead.
antagonist_name:     # Lead opposition. LoRA 2.
antagonist_refs:     # OPTIONAL on chapter 2+. On chapter 1: 15-20 references.
supporting_1_name:   # First supporting character (e.g. mentor, sidekick). LoRA 3.
supporting_1_refs:   # OPTIONAL on chapter 2+. On chapter 1: 12-15 references.
supporting_2_name:   # Second supporting character. LoRA 4.
supporting_2_refs:   # OPTIONAL on chapter 2+. On chapter 1: 12-15 references.
location_name:       # Key recurring setting (the dojo, the warehouse, the spire). LoRA 5.
location_refs:       # OPTIONAL on chapter 2+. On chapter 1: 15-20 location references.
visual_style:        # "Shōnen-Jump halftone B&W", "Seinen ink-wash", "Indie minimal line". Brand-locked.
reading_direction:   # rtl (default for manga, JP/zh-TW) / ltr (Webtoon / global / EN).
dialogue_language:   # ja (default) / en / ko / zh-tw / th / id. Loads per-locale font + SFX style.
panel_count:         # default 60. Adjust for shorter / longer chapters.
target_page_count:   # default 20 (60 panels / 3 average panels per page). Drives page-layout.
brand_kit_slug:      # kebab-case. e.g. tetsuken-saga-ch3.
chapter_slug:        # kebab-case for HTML / PDF filename. e.g. tetsuken-saga-ch3.
```

---

## Quality controls (auto-applied)

The full series quality loop, tuned for editorial-submission strictness:

- **Cost upfront.** Per-chapter plan + total cost surface before any render. Week 0 (5-LoRA training) surfaces separately.
- **Multi-LoRA entity routing.** The entity router parses each panel's prompt; when 2+ named characters co-occur, multi-LoRA blend mode picks the relevant LoRAs and weights them via the `multi-character-blocking` skill. Top-5 manga on-modelness bar in a single panel.
- **Drift gate at 0.88** (tighter than the pilot's 0.85). Editor reviews any panel below threshold; approves a re-render or accepts variation as art direction.
- **End-of-chapter cross-asset critique** — face-consistency score per character across the chapter.
- **Cross-chapter drift watch** — chapter 2+ compares each character's strongest 3 panels against prior chapter's strongest 3. Flags structural drift.
- **Auto-page-assembly with reading direction.** The `comic-page-layout` skill handles panel arrangement, gutter math, dialogue placement, SFX placement, panel-border style — all reading-direction-aware.
- **One brand-chain to three formats.** Editorial PDF + Webtoon scroll + social teaser come from one master, identically branded.

## How the agent should run this (interaction contract)

1. **CONFIRM (one message, ≤1 question):** detect the mode and restate it — "Chapter 1 = Week-0 (5 LoRAs, ~$20, ~25 min) + chapter (~$40, hours)" vs "Chapter 2+ = ~$35–50, 4–6h" — plus real choices: 60 panels vs a shorter pilot · RTL vs LTR · skip the teaser. The Step-0 STOP is this confirm; don't re-ask after it.
2. **PREVIEW CHECKPOINT:** the 60-panel beat plan (~$0.05) + one sample render per LoRA — BOTH approved before the ~$30 panel batch fires. The Step-1/3/4 STOPs are the gates; never skip them to "save time."
3. **NARRATE:** LoRA training → surface each loss curve as it lands; the 40-min panel render → a 1-liner per wave ("panels 1–12 done, drift gate clean, 2 re-fires queued"), poll `get_create_media` / `get_lora_train` every 10–15s, never silent >2 min.
4. **FAIL GRACEFULLY:** LoRA train fails → retry once → fall back to the anchor-image recipe and SAY the on-modelness drops (~85–90% vs ~95%); flux-dev panel 422/timeout → soften prompt → `director_re_render`; drift-gate trip → 3-up compare, the editor decides (never silent-retry forever); teaser `seedance-i2v` hang → `ltx-q-i2v` (`num_frames:241`) → Ken-Burns pan on the panel. ≤2 retries per panel; a chapter with 2 accepted-drift panels beats no chapter.
5. **DELIVER:** PDF + scroll + teaser + the scorecard's honest line ("cast face-consistency 0.XX; panels 14, 41 accepted as art-direction drift") + ONE next step ("run Chapter {n+1} — Week 0 is skipped, ~$35").

## How this works

Two modes. **Chapter 1** does Week-0 (5-LoRA training + brand kit + series HTML scaffold) plus the chapter. **Chapter 2+** skips Week-0 and runs only the chapter flow.

## The steps

### Step 0 — Confirm the brief + detect mode

The agent checks for an existing series with `series_slug`. If found, runs chapter-2+ mode. Otherwise Week-0 setup.

**STOP**: "Series '{series_title}' [is new / exists with N prior chapters]. Chapter {chapter_number} starting. Continue? (continue / edit brief)"

### Step 1 — Week 0 only: Train the 5-character cast (~25 min wall clock, ~$20)

Skipped on chapter 2+. The agent fires 5 `submit_lora_train` runs (one per cast member: protagonist, antagonist, supporting 1, supporting 2, location), polls each with `get_lora_train`, then `attach_lora_to_project` binds them to the chapter project. Each trigger phrase is set to the character's `_name` so the entity router can find them by name in any panel prompt. LoRAs train on the LoRA-compatible base (`flux-dev`), which is the panel-render model for the whole chapter.

> **For the agent**: surface each training job's loss curve. Once all 5 are ready, do a sample render of each character + the location to confirm look-quality before approving for the chapter run.

**STOP**: "All 5 LoRAs ready. Sample renders? (sample [character] / continue)"

### Step 2 — Week 0 only: Brand kit + finishing chain (~3 min, ~$0.10)

The agent calls `brand_kit_create` (B&W halftone screentone style enforced at the prompt-prefix level, dialogue font per `dialogue_language`, SFX font per language, panel-border style — Shōnen-Jump thick / Seinen thin / Indie minimal per `visual_style` — and reading_direction config), then `brand_kit_set_finishing_chain` to register the page-assembly → gutters → dialogue placement → SFX placement → cover → ToC → PDF + scroll + teaser steps. `set_active_brand_kit` makes it the chapter default.

**STOP**: "Brand kit + finishing chain shown. Approve, or tune one element? (approve / tune [palette/font/panel-border/screentone] [direction])"

### Step 3 — Per-chapter: 60-panel beat planning (~5 min, ~$0.05)

The agent reads `chapter_brief` against the series arc and proposes a 60-panel structure. Each panel: 1-sentence shot description + cast characters tagged inline (which LoRAs to route) + dialogue line (if any) + SFX cue (if any). Plot beats follow chapter-shape conventions for the visual_style (Shōnen escalates kinetic energy mid-chapter and climaxes panels 50-58; Seinen lingers more, climaxes earlier; Indie is freer).

> **For the agent**: chapter pacing follows the `manga-screentone` skill's panel-density rule — 2 panels per page average for emotional beats, 4 panels per page for fast action. This drives the auto-page-assembly math later.

**STOP**: "60-panel beat plan shown (~XX pages estimated). Approve, refine, or rewrite a panel? (approve / refine [#] [direction] / rewrite [#] [new]"

### Step 4 — Cost preview (~5 sec)

The agent surfaces the per-chapter cost: 60 panels × $0.50 (with multi-LoRA blend) + cover × $0.80 + page-assembly + ToC + finishing fanout = ~$40 typical. Plus cross-chapter critique pass on chapter 2+.

**STOP**: "Plan total $X.XX. Approve render? (approve / cap-at [$amount] / stop)"

### Step 5 — Render the 60 panels with multi-LoRA routing (~40 min, ~$30)

The agent runs `generate_project` (`flux-dev` base, brand kit active) across all 60 panels. The entity router auto-binds the right LoRA per character per panel from the project's attachments. Multi-character panels use blend mode with per-character spatial masking (the `multi-character-blocking` skill drives composition). Drift gate at 0.88 on every panel; `director_re_render` re-fires any panel below threshold. `critique_batch` fires after the batch for the cross-asset face-consistency score.

> **For the agent**: when the drift gate trips, surface a 3-up compare (failing panel + 2 earlier on-model panels + LoRA reference). The editor (user) makes a call: re-render at higher LoRA strength, accept as art direction, or pause to re-train the LoRA. Never silently retry forever.

**STOP**: "Panel batch + critic verdict shown. Approve all, redo specific, or accept drift on [#]? (approve / redo [#] [direction] / accept [#] / replan / pause-to-retrain)"

### Step 6 — Page assembly with dialogue + SFX (~5 min, ~$0.50)

The agent runs `comic-page-layout` on the 60 panels via the brand-kit's finishing chain. Output: ~20 pages with gutter math, panel-border style enforced, dialogue placed (clearing focal faces per `comic-typography` skill), SFX placed (per `manga-screentone` skill's onomatopoeia placement rules), reading_direction respected.

> **For the agent**: pages follow the planner's per-page panel-count (2 for emotional, 4 for action). RTL reading direction flips panel order per page; LTR keeps default. The page-layout primitive emits PNG + structured JSON (panel positions, dialogue positions, SFX positions) so future re-layouts are deterministic.

**STOP**: "Pages assembled (X pages). Approve, restructure, or re-layout specific page? (approve / restructure [direction] / re-layout [page #])"

### Step 7 — Cover + ToC (~3 min, ~$1)

The agent picks the chapter's most kinetic panel + re-renders it at editorial-cover spec: B5 hardcopy (148×210 mm @ 350dpi) + 1080×1620 digital thumbnail. Both at quality threshold 0.9 + one retry. Then generates the ToC page from chapter metadata.

**STOP**: "Cover + ToC shown. Ship, refine cover, or pick a different panel as hero? (ship / refine [direction] / re-pick [#])"

### Step 8 — Tri-format export (~3 min, ~$0.40)

The agent fires `director_export` with the chapter's brand kit. One command yields:
- **Chapter PDF** (B5 hardcopy spec, editorial-submission-ready) → `public/chapters/{chapter_slug}.pdf`
- **Webtoon single-scroll** (1080 × ~variable, Tapas / Webtoon Canvas upload spec) → `public/chapters/{chapter_slug}-scroll.jpg`
- **9:16 1-min social teaser** (strongest 8 panels animated with `seedance-i2v` for subtle motion + brand audio muxed via `ffmpeg-mux` `audio_fill:"loop"` so the bed covers the full 60s) → `public/chapters/{chapter_slug}-teaser.mp4`

**STOP**: "All three exports ready. Approve, or restart any? (approve / restart [pdf/scroll/teaser])"

### Step 9 — Chapter HTML + scorecard (~30 sec)

The agent writes `public/chapters/{chapter_slug}.html` containing cover, ToC, PDF link, scroll, teaser, dialogue script, and brand reference. Updates `public/series/{series_slug}.html` to add the chapter card. Final scorecard surfaces cast on-modelness, drift gate trip rate, and total cost vs. ceiling.

## Final output

```
✅ Chapter HTML:           public/chapters/{chapter_slug}.html
✅ Editorial PDF (B5):     public/chapters/{chapter_slug}.pdf
✅ Webtoon scroll:         public/chapters/{chapter_slug}-scroll.jpg
✅ Social teaser (60s 9:16): public/chapters/{chapter_slug}-teaser.mp4
✅ Cover (B5 + 1080×1620): <URLs>
✅ ToC page:               <URL>
✅ ~20 assembled pages:    <URL × ~20>
✅ 60 raw panels:          <URL × 60>
✅ Series HTML:            public/series/{series_slug}.html (updated)
✅ Cast LoRAs (reusable):  5 × lora_{character}_v1
✅ Brand kit (reusable):   ckit_{brand_kit_slug}

Total spent: $X.XX of $50 ceiling
LoRA training (Week 0 only): $20 one-time
Total wall-clock: HH:MM
Critic verdict: SHIP — cast face-consistency 0.XX across this chapter + 0.YY across the series
```

## What to do next

- **Submit the PDF** to a manga editorial team (Shogakukan, Shueisha, Kodansha, Image Comics) or use as the pitch artifact for an agent.
- **Upload the scroll** to Webtoon Canvas / Tapas as your monthly chapter drop.
- **Schedule the teaser** on TikTok / Reels / IG / X around the chapter drop for discovery.
- **Run this playbook again for Chapter {chapter_number + 1}** — Week 0 is skipped; cost drops to ~$35 per chapter; the cast is already on-model.
- **For an `webtoon-to-anime-trailer.md` adaptation pitch**: the 5-LoRA cast feeds the trailer pipeline directly. No re-training.
- **Year-2 cleanup**: re-train the cast LoRAs from the season's strongest 30 panels per character. v2 LoRAs encode the canonical look at peak chapter-quality.
- **For agency / publisher commissions**: the chapter PDF + series HTML are the portfolio piece. Send the URL.

---

> *"Shōnen-Jump editors don't accept a chapter because the art is good. They accept it because every character looks like themselves, every page reads in the right direction, and the climax lands on the right page. This playbook is built around exactly those three rules."* — what editorial-tier discipline looks like in production.