🎬 AI Video Agency — Magnific Edition

🇩🇪 Kurzbeschreibung

Ein Claude-Code-Skill, der aus einem Foto eine komplette filmische KI-Video-Story macht — mit durchgehend konsistentem Hauptcharakter, echtem Spannungsbogen und nahtlosen Verwandlungs-Morphs. Diese Edition läuft komplett über den Freepik/Magnific-MCP-Connector: keine CLI-Installation, kein separater Video-KI-Account, kein ffmpeg nötig. Adaption von Arnie936/ai-video-agency (ursprünglich auf Higgsfield-CLI-Basis) — die Methode (Charakter-Identitäts-Anker, Pose-Symmetrie-Morphs, keine Frame-Verkettung) stammt vom Original, hier neu implementiert auf den Magnific-Tools. Details, Änderungen gegenüber dem Original und bekannte Einschränkungen: siehe Changelog weiter unten.

A Claude Code / Claude coding-agent skill that turns one photo into a complete cinematic AI video story — consistent main character, real narrative arc, seamless transformation morphs. This edition runs entirely on the Freepik/Magnific MCP connector: no CLI to install, no separate video-AI account to sign up for, no ffmpeg on your machine.

You (photo) ──morph──▶ Character ──▶ 3 cinematic story scenes ──morph──▶ You (photo)

🙏 Source

This is an adaptation of Arnie936/ai-video-agency by Arnold Oberleiter, originally built on the Higgsfield CLI (GPT Image 2 + Seedance 2.0). All credit for the underlying method — the character-sheet identity anchor, pose-symmetry morphs, no-frame-chaining rule, story arc structure — goes to the original. This version keeps that method intact and re-implements the plumbing on top of the Freepik/Magnific MCP tools instead of a local CLI.

Original license: MIT. This adaptation is released under the same terms.

What changed vs. the original

Original (Higgsfield CLI) This edition (Magnific MCP)
Backend Higgsfield CLI, separate account required Freepik/Magnific MCP connector
Setup Install CLI, higgsfield auth login, install official skill pack None — works as soon as the connector is enabled
Identity anchor Manually prompted "character reference sheet" image Native library_create character asset (reusable numeric id)
Video model Seedance 2.0 via CLI flags Seedance 2.0 via video_generate (same model)
Final assembly Local ffmpeg concat Native video_concatenate
Camera direction Prompt text only 52 native cameraMotion presets
Reverse-morph start frame Exact last frame via ffmpeg -sseof Approximated with the Step 3 target image (known limitation, see below)

📁 Structure

ai-video-agency-magnific/
├── SKILL.md                      # the agent workflow
├── README.md
└── references/
    ├── prompting-guide.md        # proven prompt templates per phase
    └── troubleshooting.md        # identity drift, reference errors, morph fixes

🚀 Setup

Quick start (recommended)

Open a coding agent (e.g. Claude Code) in your project folder and paste this:

Clone https://github.com/ki-stuff/ai-video-agency-magnific into my skills folder (e.g. ~/.claude/skills/) and tell me what it does.

The agent installs the skill for you. There's nothing else to set up — no CLI, no separate account, no ffmpeg. Just make sure the Freepik/Magnific connector is enabled in your Claude account (Settings → Connectors).

Manual install

  1. Make sure the Freepik/Magnific connector is enabled in your Claude account (Settings → Connectors).
  2. Copy the ai-video-agency-magnific/ folder into your agent's skills directory (e.g. ~/.claude/skills/), or however your host discovers skills.
  3. That's it — no CLI, no API keys to paste, no ffmpeg.

🎯 Usage

Just tell your agent something like:

"I want to turn myself into a viking and tell an epic 30-second story. Here's my photo."

or simply:

"Make an AI video story from this photo."

The skill takes over: analysis → concept pitch → script sign-off → generation → final video.

Known limitation

The original CLI version extracts the exact last video frame with ffmpeg to use as the reverse-morph start image. Magnific has no video-frame-extraction tool, so this edition reuses the Step 3 transformation-target image instead — since the last story scene is directed to end on that same pose anyway, it's a close approximation. If Magnific ships a frame-extraction tool later, swap it back in for frame-perfect results.

💳 Costs

Generation consumes Freepik/Magnific credits on your connected account. Check account_balance before a production run — even accounts with an "unlimited" plan can still be metered per session.

Changelog

v1.0.1 — 2026-07-07

Smoke-tested end-to-end against the live API (synthetic test character, 1 morph + 1 story scene + concatenation). Two corrections from that test:

  • video_generate references do not accept the numeric library_create id — only images_generate does. Video references need an actual image creation identifier (e.g. the target image). Fixed in SKILL.md Step 5 and references/troubleshooting.md.
  • references[].type: "character" on video_generate triggered a Seedance moderation block on a photorealistic human face; type: "image" with the same identifier worked. SKILL.md now defaults to type: "image" for character consistency in video scenes.

v1.0.0 — 2026-07-07

  • Initial Magnific-MCP adaptation of Arnie936/ai-video-agency.
  • Replaced Higgsfield CLI + account with Freepik/Magnific MCP tools throughout.
  • Replaced manual character-sheet prompting with native library_create character assets.
  • Replaced local ffmpeg concatenation with native video_concatenate.
  • Added native camera-motion direction (52 presets) for story scenes.
  • Documented known limitation: reverse-morph start frame is approximated, not extracted, pending a Magnific frame-extraction tool.

📜 License

MIT — same as the original. Use it, fork it, ship videos with it.