A complete, production-tested workflow for creating realistic AI video. From idea to finished film — using Seedance 2.5, Nano Banana Pro, and Claude / ChatGPT.
This is not a list of prompts. It is a full production system — the same workflow used to make the films below, documented so anyone can replicate it.
It covers two content types that use the same pipeline:
A 30-second film about this system: why AI filmmaking looks so complex, the 7-stage workflow that fixes it, and real results made with it.
Source: /tools/launch-film/ — a single HTML file with no dependencies. The animation, particles and Web Audio soundtrack are all generated in your browser. If your browser blocks autoplay audio, tap once for sound.
These were made entirely with this workflow.
Full pipeline: character sheet → product reference → location plates → Seedance 2.5 block prompts → CapCut edit
All assets and prompts for this film are in /examples/tess-lode-ugc/
Full pipeline: brief → script → character sheets → location plates → Seedance 2.5 → After Effects
Watch the full production process explained step by step.
| Stage | Tool |
|---|---|
| Idea + Scripting | Claude / ChatGPT |
| Prompt Engineering | Claude / ChatGPT |
| Image Generation | Nano Banana Pro |
| Video Generation | Seedance 2.5 on OpenArt |
| Editing | CapCut / After Effects |
STAGE 1 IDEA
Answer 5 brief questions in plain language
STAGE 2 SCRIPT
Claude / ChatGPT — physical actions in [brackets] inline with dialogue
STAGE 3 REFERENCE IMAGE PROMPTS
Portrait → Character Sheet → Skin Realism Pass → Location Plate
STAGE 4 IMAGE GENERATION
Nano Banana Pro — in fixed order, saved with clear names
STAGE 5 VIDEO PROMPT ENGINEERING
Claude / ChatGPT — block prompts with Style DNA, timestamps, @image refs
STAGE 6 VIDEO GENERATION
Seedance 2.5, omni_reference, 1080p, native audio ON, block by block
STAGE 7 EDITING
CapCut (fast/persona) or After Effects (cinematic)
Full documentation → GUIDE.md
Describe your character in a simple form and get the two Stage 3 prompts for Nano Banana Pro, ready to copy-paste:
▶️ Open the Character Sheet Builder
Source: /tools/character-sheet-builder/ — a single HTML file, nothing to install, runs entirely in your browser.
guide-to-best-ai-filmmaking/
│
├── README.md ← you are here
├── GUIDE.md ← full workflow documentation
│
├── examples/
│ └── tess-lode-ugc/
│ ├── README.md ← breakdown of this specific film
│ ├── prompts-seedance.md ← full Seedance 2.5 block prompts A + B
│ ├── prompt-character.md ← Tess character sheet prompt
│ ├── prompt-product.md ← LODE product reference prompt
│ ├── prompt-location-kitchen.md
│ ├── prompt-location-gym.md
│ └── assets/ ← character sheet, location plates
│
├── tools/
│ ├── character-sheet-builder/
│ │ └── index.html ← Character Sheet Builder web app
│ └── launch-film/
│ └── index.html ← 30-second launch film
│
└── templates/
├── film-brief-template.md ← blank brief for cinematic films
├── persona-brief-template.md ← blank brief for UGC / persona content
├── character-sheet-prompt.md ← fill-in character sheet prompt
├── location-plate-prompt.md ← fill-in location plate prompt
└── seedance-block-template.md ← fill-in block prompt for Seedance 2.5
Name the mechanism, never the mood. Not “make it cinematic” — name the lens, the light source, the color grade. Not “make it realistic” — say it was shot on a real camera, not a 3D render. Not “she looks scared” — describe what a camera would actually see.
Lock character first, location second. Every later prompt says “match @Image1” instead of re-describing from scratch. A model inventing a location and lighting at the same time as a performance will fail both.
Physical actions only. Never interior states. “Her jaw tightens, she doesn’t blink” — not “she is nervous.” A video model renders what a camera sees. Give it something to see.
Style DNA is written once, pasted byte-identical. Never reword it between blocks. One project, one signature.
Dialogue and physical action never overlap. She speaks, or she moves. Never both at the same time.
GUIDE.md — the full workflow/templates/film-brief-template.md or /templates/persona-brief-template.md/examples/tess-lode-ugc/ for a real completed exampleIf you use this workflow and make something you’re proud of, open a pull request and add your example to /examples/. Include:
The more real examples this repo has, the more useful it becomes for everyone.
MIT — use freely, credit appreciated.
Made by Fazi — AI filmmaker and multi-language YouTube creator. YouTube