Our launch video is a 1985 office on the day the PCs arrive and nobody's trained anyone yet. This guide is the how behind it: the actual steps, in order, plus the one mistake that cost me the most rework.
Claude (concept + coaching) → ChatGPT (reference images + prompts) → Higgsfield + Seedance 2.0 (video generation) → DaVinci Resolve (edit)
Swap in whatever tools you already use. The process holds either way.
Define the concept
Your brain, or your brain plus AI, either way works, but the idea itself has to come from you. AI isn't where a concept starts.
Challenge it with AI
Give your model the goal, the channel, and the audience, then have it push back until the direction actually holds up. This is also where the practical calls get made: vertical or square, target length, voiceover or none. Don't lock those in before you've got a direction to test them against.
Write the beat sheet
Scenes and beats, in text, no images yet. It's a storyboard, just written instead of drawn.
Align visually with a handful of reference images
Four or five images, just to lock down characters and environments. Not lighting, not every shot: just enough that you know what things look like.
Build a prompt generator
Ask your LLM to help you set up a skill that turns each scene into the right prompt for your video-generation model. Feed it your refined concept and beat sheet, so it's built for this project specifically, not a generic prompt-writing assistant.
Pick a platform and model
We went with Higgsfield + Seedance 2.0. Use whatever fits your project. Whichever you choose, check the credit cost per generation before you start, not after.
7 · Test ONE scene before generating everything
If something in your process is off, you want to find out from one lost scene, not from fifteen.
8 · Generate prompts scene by scene
Give your prompt generator a written, directional description of your references, not the images themselves. Feeding images straight to the video model over-constrains it: it locks onto reproducing that exact frame instead of directing the shot, and you lose the execution. Words pin down what has to stay true (character, environment, period) and leave the how open. Over-constraining was my most expensive mistake.
Save characters and environments as assets
As they get generated, save them on the platform so you can reuse them and keep continuity across scenes.
Stitch the shots
Two paths, depending on how much control you want.
DaVinci Resolve (free, and the more precise editor) with your LLM as a coach for captions, color, and audio mix.
An agent with file access, Fable (Claude) or Astra (ChatGPT), stitches the shots for you: it looks at each clip, asks you for the order, and assembles the final cut.
Nothing here is hard, it's just unfamiliar. The generation step took the least time. Writing a concept that survives being challenged by AI, and describing your world precisely enough in words that a model never even sees your references, took the most. That's where the real work is.
Curious about applying AI in your own organization? Message me on LinkedIn, happy to talk it through.