October 2026
How I write a video prompt that actually holds together
A video prompt isn't a wish list. It's a short direction written for something that takes you literally. Here's the structure I settled on after the RESILIENT video.
Start with the frame, not the list
I open with what the camera sees and how it moves — the shot type, the motion, and the timing of that motion. Everything else hangs off the camera because that's what the viewer actually perceives first.
If I start by listing props and colors, the model spends its attention on the wrong layer and the motion comes out vague.
Write timing in numbers, not adjectives
'Slow motion around the fall' is ambiguous. 'The fall happens over 4 seconds, then a 2-second freeze on the airborne objects' is something the model can follow.
I split the runtime into beats and give each beat a duration. The RESILIENT prompt is structured this way — that's why the airborne moment landed at all.
Sound direction belongs in the prompt
In Higgsfield with Seedance 2.5 the sound is guided inside the same workflow, so I describe it in the prompt: the source, the moment it enters, and the energy curve.
Separating sound into a later step breaks the pacing — the cut follows the sound, so the sound has to be specified up front.
Shorter and structured beats longer
My early prompts were long because longer felt safer. The model doesn't read length the way a person does — it follows structure. A tight prompt with clear sections outperforms a wall of detail.
When I cut the RESILIENT prompt back to its structure, the results got more consistent, not less.
Keep reading
- The full RESILIENT tutorial with the actual prompt
- My projects and experiments
- Reference images: the part most people skip
- Behind the RESILIENT video: what the tutorial doesn't show
- The AI tools I'm testing this month
- What I learned shipping AI content publicly
- My creator routine in Dubai