reelforge

Turn one topic into a production-ready short-form video pack — a locked story, then, for every scene, a Step 1 image prompt and a Step 2 video prompt, serialized into each model's official native format.

reelforge is an agentic (LangGraph + Claude) engine. It plans the story, locks a Story Bible so characters never drift, and emits an image-first production pack: a still for each scene, then the image-to-video prompt that animates that exact still — chained scene to scene.

Why this design? A one-shot "make me a video prompt" script is easy to copy. A consistency-locked, per-scene, image-first pipeline that serializes to the real Nano Banana Pro / Seedance 2.0 grammar is the part that's hard to replicate — and it's deliberately the part kept in the open here. The premium prompt library (the tuning) stays private. See Safe by design.

---

Architecture

flowchart LR
    T([Topic]) --> P[Planner<br/>logline · hook · length]
    P --> B[Story Bible<br/>lock characters · props · look]
    B --> S[Scriptwriter<br/>hook → turn → cliffhanger]
    S --> A[Assembler<br/>per-scene, image-first]
    A --> O([Content Pack])

    subgraph perscene [For every scene]
      direction LR
      I["Step 1 · IMAGE<br/>(Nano Banana Pro)"] --> V["Step 2 · VIDEO<br/>(Seedance 2.0)"]
    end
    A -.-> perscene

    classDef agent fill:#1e293b,stroke:#7C3AED,color:#e2e8f0;
    class P,B,S,A agent;

Loading

Stage	Model / role	Output
Planner	Claude	logline, theme, audience, scroll-stopping hook, target length
Story Bible	Claude	locked characters (+ Soul ID), locations, props, look — the anti-drift contract
Scriptwriter	Claude	beat-arc storyboard (hook → setup → turn → rock bottom → cliffhanger)
Assembler	pure code	per scene: Step 1 image prompt + Step 2 video prompt, in native format

---

Per-scene, image-first — the core idea

For every scene the engine emits two steps, in order:

1. Step 1 — IMAGE: a prompt that makes the scene's still (keyframe). Built from the Story Bible and locked to an anchor hero still via @image1, so the face, hair and wardrobe are identical scene after scene.
2. Step 2 — VIDEO: an image-to-video prompt that animates that still — motion and one camera move only, never re-describing the image — and chains from the previous clip (Extend @video1) so the whole story flows.

Consistency comes from the Bible (pulled verbatim into every Step 1) + the anchor reference; story logic comes from beat order + clip chaining.

Native-format serializers (the craft)

The same structured scene serializes into each model's official grammar:

Step	Model	Native format
Image	Nano Banana Pro (default)	natural-language 5-pillar, @image1 identity lock, never JSON
Image	Soul Cinema	Soul ID + tight prose + real camera/lens
Video	Seedance 2.0 (default)	I2V: motion + one camera move, positive constraints, Total: Ns / 1 shot / 9:16 footer
Video	Kling 3.0	I2V with a real negative-prompt field

Image and video each have two interchangeable targets — pick one per run with --image-model / --video-model (defaults in bold).

---

Quickstart

# Python 3.10+
pip install -e ".[dev]"
cp .env.example .env          # add ANTHROPIC_API_KEY

reelforge "A barista discovers the cafe is laundering money" --scenes 5 -o out/pack.json

No key yet? The pipeline still proves itself — the serializers and the per-scene assembly are pure and deterministic:

python examples/build_sample.py     # writes examples/sample_pack.{json,md} — no API needed

See examples/sample_pack.md for a full generated pack (a 5-scene betrayal story: anchor still → per-scene Step 1/Step 2 → chained clips).

Choose models per run:

reelforge "topic..." --image-model soul_cinema --video-model kling_3_0

---

Safe by design

This is a public portfolio repo, built so the work can't simply be lifted:

• The engine is open; the tuning is not. prompts/sample/ ships generic templates so it runs end-to-end. The production library (viral-structure frameworks, per-niche tuning, persona/brand-voice configs) is proprietary and loaded only from a private REELFORGE_PROMPTS_DIR. See prompts/README.md.
• Restrictive license. Source-available under PolyForm Noncommercial 1.0.0 — look, learn, evaluate; no commercial use or resale.
• Provenance. Public commit history under the author's name is the record that this work is original and whose it is.

---

Project layout

src/reelforge/
  config.py       # env-driven settings (nothing secret hard-coded)
  schema.py       # Pydantic contract: ContentPack / Scene / StoryBible / Image+Video prompts
  llm.py          # Claude helper — adaptive thinking + robust JSON extraction
  prompts.py      # prompt loader (private dir → sample → in-code default)
  serializers.py  # Scene → native prompt (Nano Banana Pro / Soul Cinema / Seedance / Kling)
  state.py        # LangGraph state
  nodes.py        # plan / story bible / script / assemble
  graph.py        # graph assembly + run()
  render.py       # ContentPack → readable markdown
  cli.py          # `reelforge "<topic>"`
prompts/sample/   # generic, shippable templates (the moat stays private)
examples/         # build_sample.py + a committed, no-API content pack
tests/            # pure serializer + schema tests (run in CI)

License

Source-available under PolyForm Noncommercial 1.0.0. Copyright © Daniel Enache.