One creative brief goes in. A full set of production-ready image prompts and cinematic video prompts comes out, ready to paste into any generator you already use. The whole skill is below. Setup takes five minutes.
Most AI video looks like AI video. You can tell in the first second. Faces drift between shots. The lighting has no logic. The camera sits still the way no real camera ever does.
The model is rarely the problem. The prompt is. A one-line brief gives the generator nothing to hold onto, so it fills the gaps with mush.
This skill writes the prompt for you. Not one prompt. A whole layered set: character sheets, scene layouts, and cinematic video prompts with camera and lens already decided.
What The Skill Builds
It works in two layers.
The first is the image layer. It reads your brief and writes a character sheet for every person who appears: physical description, outfit, expression, and an image prompt tuned for consistency across shots. It does the same for every scene: location, lighting, composition, and a prompt specific enough that any image model returns the same look twice.
The second is the video layer. It takes each scene's image prompt and expands it into a cinematic video prompt. Camera movement, lens choice, shooting style, clip duration. All decided, all written, ready to paste into your video tool.
Brief in the top. Production-grade prompts out the bottom. You still run the images and the clips yourself, so you stay in control of every frame.
Copy The Skill
Save this as SKILL.md in a skills/ai-video-pipeline/ folder, or paste the whole thing into a new chat and tell your agent to follow it.
Copy this.
---
name: ai-video-pipeline
description: Build a complete AI video prompt set from a creative brief — generates a production-ready image layer (character sheets, scene layouts, detailed image prompts) and a cinematic video prompt set (camera movements, lens settings, shooting styles). Use when you want studio-quality AI video output from a text brief.
---
# AI Video Production Pipeline
> What this does (in one line): Feed a creative brief in, get a full set of production-ready image prompts and cinematic video prompts back — ready to send to any AI video generator.
## BEFORE YOU START (fill these in)
These are the only things you need to provide. Everything else is automatic.
| Fill in | What it is | Example |
|---|---|---|
| BRIEF_FILE | Path to your creative brief — describe your concept, characters, scenes, mood, and visual style | ~/projects/my-video-brief.md |
| OUTPUT_FOLDER | Where to save the two output files (image-prompts.md and video-prompts.md) | ~/projects/video-pipeline/ |
> If you do not have one of these yet, the skill will tell you exactly where to get it the first time you run it.
## What you need installed
- Claude Code (or Claude Desktop with file access)
- A plain-text brief file describing your video concept (see "How to write a brief" below)
- Any AI image generator (Midjourney, Flux, DALL-E, Ideogram) for Step 2
- Any AI video generator (Kling, Runway, Sora, Hailuo, Pika) for Step 3
## How to use it
1. Put this SKILL.md where your agent looks for skills (in Claude Code: a skills/ai-video-pipeline/ folder, or paste the whole file into a new chat and say "follow this skill").
2. Fill in the fields above.
3. Then just tell your agent: "Build the complete image and video prompt set from my brief at [path to brief]"
## How to write a brief
Your brief does not need to be long. One to two paragraphs covering:
- Characters: Who appears. Physical description, outfit, age range, mood.
- Scenes: Where the action takes place. List each key scene or beat.
- Visual style: How it should look. Reference a film, a photographer, a color palette, or a lighting style.
- Tone/mood: What it should feel like. Fast-paced, contemplative, cinematic, raw.
Short example:
Character: A man in his 30s, dark suit, confident posture, walking through a
glass-walled office at night. The city glows outside.
Scene 1: He enters the office, sets a laptop on a desk, opens it. Blue light
hits his face.
Scene 2: Close on the screen — code running, charts updating.
Scene 3: He leans back, exhales. Outside the window, the city. Wide shot.
Style: David Fincher. Cold blue-white light. High contrast. Cinematic 2.39:1.
Tone: Tense but controlled. Late-night focus.
## The workflow (what the agent does)
Step 1 — Read and parse the brief
Load BRIEF_FILE. Extract: characters (name, visual description, outfit, expression), scenes (location, action, beats), visual style, mood, and any specific requirements noted.
Step 2 — Build the Image Layer
For each character: write a character sheet with full visual description, outfit details, and a detailed image generation prompt optimized for consistency across shots.
For each scene: write a scene layout covering location, lighting setup, composition (foreground/mid/background), and a detailed image generation prompt. Prompts should be specific enough that any image model produces a consistent result.
Save all of this as image-prompts.md in OUTPUT_FOLDER.
Step 3 — Build the Cinematic Video Prompt Layer
Take each scene's image prompt and expand it into a cinematic video prompt by adding:
- Camera movement: static / slow push in / dolly left / crane up / handheld / rack focus
- Lens: wide (24mm) / medium (50mm) / close-up (85mm) / extreme close-up
- Shooting style: name a specific visual reference or lighting condition
- Duration: suggested clip length in seconds
Save as video-prompts.md in OUTPUT_FOLDER. Each entry carries the base image prompt plus the camera, lens, style, and duration combined into one full prompt ready to paste into the video tool.
Step 4 — Summary
Tell the user: how many image prompts were built, how many video prompts were built, and where the files are saved. Flag any scenes where the brief was thin and the prompts may need manual refinement.
## If something goes wrong
- "Prompts feel generic" -> Add more specific visual references to your brief. Name a director, a film, or a photographer. Generic briefs produce generic prompts.
- "Video model ignores camera movement" -> Put the camera instruction at the very start of the full prompt, not the end. Most models weight the opening tokens heavily.
- "Output folder does not exist" -> The agent will create it. Provide the full path, not a relative one.
- "Characters look inconsistent across scenes" -> Use the exact character-sheet prompt verbatim in every scene that features that character. Do not paraphrase it per scene.
How To Use It
Write a brief and save it as a plain text file. Fill in the two fields, BRIEF_FILE and OUTPUT_FOLDER, at the top of the skill. Then tell your agent to build the complete image and video prompt set from your brief, and read the two files it hands back.
Done. You now have a full prompt set instead of a blank generator box.
What You Get Back
01 · Character sheets that hold. Same face, same outfit, every shot, because the sheet is written once and reused verbatim.
02 · Scene layouts. Location, lighting, and composition worked out before you spend a single generation credit.
03 · Image prompts you can reproduce. Specific enough to return the same look twice, not a one-off lucky render.
04 · Cinematic video prompts. Camera movement, lens, shooting style, and duration already decided and written into the prompt.
Honest: What This Does Not Do
This does not generate the images or the video. It writes the prompts. You still run them through your own image and video tools, and you still pay those tools' bills. That is on purpose. You see every frame before it goes anywhere.
It is only as good as your brief. A vague brief gives you vague prompts. Name a director, a film, a lighting style. Two sentences of real reference beat a page of adjectives.
Consistency has a ceiling. Image models still drift between shots. Paste the exact character-sheet prompt into every scene, and expect to re-roll a shot or two.
The One Thing To Do This Week
Write a two-paragraph brief for a video you actually want. One character, three scenes, one visual reference. Then run the skill on it and read the prompts it gives back.
You will see the gap between what you were typing before and what a real production brief looks like. Worst case, you spent ten minutes and learned how the prompt is supposed to read.