Skip to content
Projects

mengxiujie/ManjuForge has 32 GitHub stars and 5 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-15 05:38 UTC. Last push 2026-10-13.

How these counts are measured

Measured star growth

in 24h
0
in 7 days
Collecting history
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

AI 漫剧/短剧连载创作引擎(Agent Skill)| 一句话到剧本+分镜+配音+AI提示词的工业化流水线 | 七个记忆文件+22项跨文件一致性检查,解决长线连载人设崩坏与伏笔丢失 | Agent-agnostic (Claude Code / Codex / Copilot / Hermes) · MIT-0

AI

README_EN.md

Original

ManjuForge — AI Manju Serial Production Engine

An Agent Skill for serialized AI manju (comic-drama) creation — works with Hermes Agent, Claude Code, Codex, GitHub Copilot, and any Agent Skills–compatible runtime.

Others make a beautiful single episode. We make an IP that survives 30+ episodes.

GitHub | 中文

Version License Platform


What is this?

ManjuForge is a serialization-focused AI creation engine for manju (漫剧, comic-drama short series popular on Douyin-style vertical platforms). It turns a one-line concept into production-ready deliverables per episode:

Full script + storyboard (16:9 horizontal / 9:16 vertical AI-execution) + voiceover copy + AI image/video prompts.

It solves the problem every 30+ episode series hits: character drift, lost foreshadowing, world-building inconsistency, and relationship confusion.

Positioning: visual consistency vs narrative memory

Most 2026 AI short-drama tools solve visual consistency (same character/scene across shots). ManjuForge solves narrative consistency across episodes:

Asset-based workspaces (Jellyfish, ArcReel, ...) ManjuForge
What stays consistent Visual assets: character/scene/prop/costume appearance Narrative memory: personality, relationships, foreshadowing, timeline, world rules
Prevents Shot-to-shot drift Cross-episode collapse (OOC at E30, dropped setups, drifting settings)
Timescale shot → episode episode → whole series (evolving visual motifs, trust curves, relationship stages)
Mechanism asset library + reference images 7 memory files + 3 engines (Story Bible / Character Graph / Foreshadow Engine)

They keep the picture stable; we keep the story stable. Complementary — ManjuForge's prompt output feeds any generation workspace.

Proven in production

  • Demo 1: 《代码赘婿》6 episodes — 0 character breaks; 100% foreshadowing closure (11/11); dual mainlines from E4; personal → faction conflict by E6. Third-party rating: A / A+ across six dimensions (Character Memory, Relationship Graph, State Engine, Economic Logic, World Consistency, Narrative Engine).
  • Demo 2: 《渊烬》Unit 1 (E1–5) — greenfield run (not retrospective): quality curve E1 4 fixes → E2 1 → E3–5 zero; mean 91.1/100 per episode; unit rating 6×A+ / 1×A.

Core capabilities

  • 7 memory files: character files / worldbuilding / scene layout (with lighting state) / plot summary / foreshadowing ledger / story bible (refreshed every 5 eps) / character relationship graph
  • 3 engines: Story Bible (single source of truth), Character Graph (relationship matrix + event timeline with trust curve), Foreshadow Engine (setup → ferment → advance → payoff lifecycle)
  • 18-point automated quality check: python scripts/quality_check.py <episode file> [--villain X --foreshadow ... --bible ... --graph ...] — 🔴 items block advancing to the next episode
  • Knowledge base: 29 reference files + 13 film-teardown case studies + 6 style templates + hook formula library (6 openings / 4 cliffhangers) + psychological trigger system
  • Vendor-decoupled execution contract (v1.8.0): prompts never hard-bind a generator; any image/video production goes through an adapter layer with a 4-step confirmation gate (bounded job → preview → explicit confirm → run), preventing runaway cost. See references/adapter-contract.md.

Install

The install directory must be lowercase manjuforge (Agent Skills spec: SKILL.md name must match its parent directory).

Hermes Agent

git clone https://github.com/mengxiujie/ManjuForge.git ~/.hermes/skills/creative/manjuforge

Claude Code

git clone https://github.com/mengxiujie/ManjuForge.git && cd ManjuForge
mkdir -p ~/.claude/skills && ln -s "$PWD" ~/.claude/skills/manjuforge

Or just tell Claude Code: install these skills https://github.com/mengxiujie/ManjuForge

Codex (OpenAI)

git clone https://github.com/mengxiujie/ManjuForge.git && cd ManjuForge
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" && ln -s "$PWD" "${CODEX_HOME:-$HOME/.codex}/skills/manjuforge"

GitHub Copilot CLI / cross-client: copy to ~/.copilot/skills/manjuforge, .github/skills/manjuforge/ (project), or ~/.agents/skills/manjuforge (cross-client standard).

China mirror: git clone https://gitcode.com/gcw_LRlig9U4/manjuforge.git <same target> if GitHub is slow.

Structure

ManjuForge/                        # skill root (dir name = skill name)
├── SKILL.md                       # core definition (frontmatter: name/description/license/metadata)
├── skill.json                     # metadata
├── prompt.core.txt                # core engine rules (open, stable)
├── LICENSE                        # MIT-0
├── scripts/quality_check.py       # 18-point automated checks
├── templates/                     # 6 memory templates
├── references/                    # 29 reference files + knowledge_base/
├── demo/                          # 6-episode complete demo
└── docs/

License

MIT-0 — free for commercial use, no attribution required, closed-source modification allowed.

Contribute

Star, fork, and open issues welcome: https://github.com/mengxiujie/ManjuForge/issues