AI Creator Video Engine: Consistent AI Video, Any Style

What it does

Takes a topic to a finished AI-generated video, in whatever form you want. It plans the video, locks the look, writes and checks the speech, compiles copy-paste prompts for Google Flow, Veo, Kling, Runway, Sora, Hailuo or any image model, and then assembles the clips you generate into a finished video: cross-dissolves, voiceover, motion infographics, karaoke captions and a music bed.

Any kind of video

• AI presenter or avatar talking to camera, with a locked face, wardrobe and studio across every cut
• Faceless video: documentary, b-roll, explainers, listicles, product films and still-image slideshows under a voiceover
• Animation: 2D, 3D, anime, claymation or motion graphics, with recurring characters who can speak in the clip or be narrated
• Mixed: a presenter who cuts to b-roll, or an animation with a narrator

Your format, your call

Every script reads your decisions from one manifest:

• Mode: presenter, voiceover or animation, set per video or per scene
• Length: from a 10-second ad to long-form YouTube planned in chapters
• Scene length: 3 to 15 seconds; the speech budget is derived per scene (1.9 to 2.3 words per second)
• Aspect ratio: 9:16, 16:9, 1:1, 4:5 or 3:4, prompted and rendered correctly, with no stretched sources
• Shots: A-Cam and B-Cam for presenters, or any shot list (wide, macro, aerial, top-down, POV, your own)
• Also yours to set: infographic cap, dissolve length, frame rate, captions, music and voiceover levels

How it works

1. Lock the look. A character sheet for a presenter, or a style bible for faceless footage and animation: medium, palette, lighting, lens or line, texture, plus fixed anchors for every recurring character, product and location.
2. Storyboard. Beat structures for explainers, documentaries, listicles, story arcs, kids' lessons, ads and long-form chapters.
3. Validate. validate_manifest.py checks every scene's dialogue or narration against its length. It also checks shot variety and the infographic cap, and flags any request for rendered text, which AI models garble. The exit code tells your agent when the manifest is clean.
4. Compile prompts. compile_prompts.py writes one prompt per scene. Presenter scenes get a voice block, a visual block and a negative shield. Faceless and animated scenes get a style lock, the shot and camera move, subject anchors and the negative shield. Narration goes to a separate script for any text-to-speech tool or voice artist.
5. Assemble. assemble_episode.py handles everything after generation.
   • Turns stills into moving shots and fits every source to the frame.
   • Gives silent clips a silent track, so picture and sound stay in step.
   • Stitches with cross-dissolves and overlays the infographics.
   • Lays the voiceover on each scene, lowering background sound but never in-clip dialogue.
   • Burns Whisper karaoke captions, mixes the music bed, and exports the video plus a thumbnail.

What's inside

• SKILL.md workflow covering every mode, the format decision table, the consistency rules and speech budgets by scene length
• Eight references:
  • the consistency rules
  • the presenter 3-block prompt formula
  • the visual prompt formula for faceless and animated scenes
  • the character sheet and style bible templates
  • storyboard beat structures for every mode
  • publishing metadata for YouTube, Shorts, TikTok, Reels and LinkedIn
  • the post-production pipeline guide
• Working examples: a presenter episode, a faceless coffee documentary, and an animated short with a speaking robot character
• Six Python scripts: the validator and prompt compiler (standard library only) and the FFmpeg and Whisper pipeline

Where it stops

It does not generate clips, images or voice audio itself. Your video, image and text-to-speech tools do that from the prompts and the narration script. The skill then assembles the results. The pipeline needs FFmpeg on your machine and, for captions, the openai-whisper package. Speech budgets are calibrated for English.

Licence and updates

One purchase gives you a personal licence to use the skill with your own agents for your own channels, projects and clients. Updates ship through Agensi and reach your agent automatically.

Keywords

ai video, faceless video, ai animation, ai avatar, character consistency, style consistency, b-roll, voiceover, storyboard, google flow

Use cases

• Plan and script an AI presenter explainer, a faceless documentary, or an animated short from one topic
• Lock a presenter, an animated character, a product or a documentary look so every shot matches
• Turn a storyboard into copy-paste prompts for Google Flow, Veo, Kling, Runway, Sora or an image model
• Write a narration script timed to each scene for a text-to-speech tool or voice artist
• Check that every scene's dialogue or narration fits its length with no cutoff or dead air
• Assemble generated clips and stills into a finished video with dissolves, voiceover, captions and music
• Plan a long-form YouTube video in chapters, or a 15-second ad, with the same guardrails
• Write YouTube titles, descriptions and chapters, or Shorts, TikTok, Reels and LinkedIn captions

Known limitations

Does not generate clips, images or voice audio itself. It prepares validated prompts and a narration script for your own tools, then assembles what they produce. The assembly step needs FFmpeg locally and, for captions, openai-whisper (a model download on first run). Speech budgets are calibrated for English; other languages can pass fixed --min and --max values. Scene lengths beyond 15 seconds exceed what most current video models hold reliably.

Compatibility note

Works with any agent that reads SKILL.md: Claude Code, Codex CLI, Cursor, Gemini CLI, OpenClaw and others. Python 3.8 or newer. FFmpeg is needed only for the assembly step. The planning, validation and prompt steps have no dependencies at all.

FAQ

Does it only make talking-head videos? No. It has three modes: an AI presenter talking to camera, faceless video (documentary, b-roll, explainers, listicles, product films, still slideshows) under a voiceover, and animation (2D, 3D, anime, claymation, motion graphics) with characters who can speak or be narrated. Scenes can mix modes.

How does it keep characters and style consistent? A character sheet or style bible fixes the medium, the look (palette, lighting, lens or line, texture, rendering) and three to five anchor phrases for every recurring character, product and location. The compiler repeats those exact words in every prompt, and the negative shield bans the usual failure modes for that style.

Does it generate the video or the voiceover? No. You generate clips or stills in your own video or image model and the voiceover in your text-to-speech tool, from the prompts and narration script the skill writes. The bundled pipeline then assembles everything into the finished video.

Do I have to use 9 scenes of 10 seconds in 9:16? No. That is only the presenter default. Scene count, scene length (3 to 15 seconds), aspect ratio (9:16, 16:9, 1:1, 4:5, 3:4), shot plan, infographic cap and dissolve length are all set in the manifest.

Why limit words per scene? AI speech and narration run at roughly 114 to 138 words per minute. Too many words get cut off, rushed, or spill into the next scene; too few leave dead air. The validator derives each scene's budget from its length and reports the pace.

Can I use still images instead of video clips? Yes. Name a scene file scene_03.png instead of scene_03.mp4 and the assembler turns it into a moving shot with a slow push-in at the scene's length, filling the frame without stretching.

What does the licence allow? One purchase gives you a personal, non-transferable licence to use the skill with your own AI agents and to produce videos for your own channels, projects and clients. You keep everything you make with it. Redistributing or reselling the skill itself is not allowed.

What do I need installed? Python 3.8 or newer. For assembly, FFmpeg on your PATH or pip install imageio-ffmpeg. For captions, pip install openai-whisper. The planning, validation and prompt steps need nothing else.

Which agents does it work with? Any agent that reads SKILL.md, including Claude Code, Codex CLI, Cursor, Gemini CLI and OpenClaw.

Do I get updates? Yes. New versions ship through Agensi and reach your connected agent automatically, with a changelog.