ArcReel: Multi-Agent Pipeline Turns Written Stories into Short Videos

By Prahlad Menon 3 min read

Most AI video tools skip the hard part: keeping characters looking the same across multiple shots. You generate a beautiful first frame, extend it into video, and suddenly your protagonist has a different face, different clothes, different everything.

ArcReel solves this by treating character consistency as a first-class concern—not an afterthought.

The Architecture That Matters

ArcReel is built on the Claude Agent SDK using an “orchestration skill + focused subagent” multi-agent architecture. Here’s the pipeline:

  1. Novel parsing → Extract characters, relationships, and timeline into a queryable graph
  2. Character design generation → AI creates reference images for each character FIRST
  3. Episode planning → Automatic chapter segmentation and beat planning
  4. Script generation → Multiple modes (adaptive, literal, staged) with review loops
  5. Storyboard generation → All images reference the character designs from step 2
  6. Video synthesis → Veo 3.1, Seedance, Grok, Sora 2—your choice
  7. FFmpeg composition → Final assembly with subtitles

The key insight: by generating character design sheets upfront and using them as references throughout the pipeline, every storyboard and video frame maintains visual consistency. Props and locations (“clues” in their terminology) get the same treatment.

Multi-Model, Multi-Vendor

ArcReel is provider-agnostic:

Image generation: Gemini (Nano Banana 2), Grok, OpenAI (GPT Image 2), Vidu, MiniMax, Kling, custom OpenAI-compatible endpoints

Video generation: Veo 3.1, Seedance 2.0, Grok, Sora 2, Vidu Q3, and others

Text/reasoning: Gemini 3.1 Pro, GPT-5.5, Grok 4.20, Qwen—whatever you configure

Switch providers at the project or global level. The abstraction layer handles the differences.

Three Content Modes

  • Narration mode — Split by reading rhythm for audiobook-style content
  • Drama mode — Structured by scenes and dialogue for short episodic series
  • Ad/short mode — Target duration, product shots, single video output

The ad mode is particularly interesting for e-commerce—upload product photos, generate a standard 8-beat promotional script, keep the product visually anchored throughout.

The Agent Architecture

The main agent doesn’t try to do everything. It delegates:

Main Agent → manga-workflow (orchestration skill)
                ↓
    ├── analyze-characters-clues (subagent)
    ├── split-narration-segments (subagent)
    ├── normalize-drama-script (subagent)
    ├── create-episode-script (subagent)
    └── asset-generation (subagent)

Each subagent completes one task and returns a summary. The heavy context (full novel text, etc.) stays inside the subagent—the main agent only sees the distilled output. This protects context windows and keeps the orchestration clean.

Crucially, there’s a confirmation step between phases. The agent presents results, waits for user approval, then proceeds. No runaway generation.

The Cost Reality

API costs add up quickly. ArcReel tracks expenses across all providers and gives you cost estimates before generation—at the project, episode, and shot level. You can see estimated vs. actual costs and make informed decisions about which provider to use for what.

The self-hosted option (AGPL-3.0 license) means you’re not locked into any SaaS pricing.

Running It

git clone https://github.com/ArcReel/ArcReel.git
cd ArcReel/deploy
cp .env.example .env
docker compose up -d
# Access at http://localhost:1241

Requires Python 3.12+, React 19 frontend, FastAPI backend. No GPU needed locally—all inference runs through the configured provider APIs.

Why This Matters

The character consistency problem has been the gap between “cool AI video demos” and “actually usable for production.” Kling, Seedance, Veo—they’re all impressive frame-to-frame, but maintaining identity across a multi-shot narrative has required manual reference image management and prompt engineering.

ArcReel bakes that workflow into the tool itself. The pipeline enforces consistency because the architecture demands it.

Still early days, and the API costs for a full short video project aren’t trivial. But the architecture is the right one—multi-agent orchestration with character identity as a system-level concern, not a prompt hack.

GitHub: github.com/ArcReel/ArcReel

If you’re building story-to-video workflows and tired of characters shapeshifting between shots, this is worth exploring.