Recently I put Seedance 2.5 and Wan 3.0 head to head. The two best video models out right now.
And my DMs filled up with the same question.
"How do you actually prompt this?"
People have seen clips of what these models can do,
yet still open the prompt box and freeze.
I used to be like that.
I test every new AI tool the week it drops. That is the job.
And for a long time every new model meant starting over, guessing at what prompts would work best for that specific model.
Then I found a highly ignored shortcut!
It takes five minutes to set up and lets you never have to manually prompt again.
What you need:
Claude/ChatGPT/Gemini - Any AI that supports skills.md.
5 minutes
Step 1. Find the documentation
Every serious model has one. Search online for “[Model] Prompt Guide/User guide“
Or go to the official website of the model, scroll to the bottom, and look for "User Guide", "Docs", or "Model Prompt Guide".
That page is the cheat sheet. It tells you what the model was trained to respond to. Camera terms it understands. How it handles motion. How long a prompt should be. What breaks it.
Almost nobody opens it. That is why most prompts are guesswork.
To save you time: Wan 3.0 & Seedance 2.5 Docs
Step 2. Copy & Paste
Select everything on the page. Every section, every example prompt, every rule.
Copy it all, and paste it into your AI of choice, together with this prompt:
This is the official user guide for [Insert Model Name]. Read all of it. Then turn it into a skill I can reuse every time I want to prompt this model.
The skill should include:
What the model does best and what it struggles with
The exact prompt structure it responds to, with the order of elements
The camera, motion, and style terms it understands
Prompt length and formatting rules
3 example prompts written in the ideal format
A short checklist to run before generating
When I give you a rough idea, use this skill to write a finished prompt for [Insert Model Name].
With this skill now made and saved. Every rough idea you give it comes back as a prompt written the way that model actually wants.
Why this beats "learning to prompt"
I see posts like this daily. "Opus 5.5 vs GPT 6, which one makes the better video."
The funny part: Neither model makes videos.
What that test really shows is which raw model wrote the better prompt from the same input.
Truthfully the raw model has no idea how Seedance or Wan actually wants to be prompted. It is guessing, just with better words than you.
But when given the docs, even a way cheaper model with the skill will beat a top model without it. And a top model with the skill just gets you the best output there is, every time.
How I use it now
Wan 3.0 has a skill. Seedance 2.5 has a skill. Every video model I use has one.
I get a concept idea. I describe it in two lines and ask Claude to write it with the Seedance skill.
I find a prompt concept online I like. I paste it in as a reference, tweak it to my liking, and have it rebuilt for the video model I am using.
A new model drops - like Kling 4.0. Docs in, skill out, full speed on day one.
As models keep getting better, this framework stays the same.
And it’ll help you explore your ideas with gen AI without the technical side being your main constrain.
Catch ya in the next one!
Best,
Melvin
Melvin Hagström
[email protected] | @ai.volve.news
