September 2026

Behind the RESILIENT video: what the tutorial doesn't show

The tutorial gives you the prompts and the reference images. This note is the part that doesn't fit in a step list: the thinking, the retries, and the choices I'd make differently.

The character sheet decided everything

Before writing a single line of the video prompt, I built a character reference sheet from three photos: a selfie, the outfit, and the sunglasses. Every later attempt stayed consistent because of that one image.

When I skipped this step in earlier experiments, the face changed between shots and no amount of prompt detail fixed it. The reference image did what a paragraph couldn't.

What took the most attempts

The airborne moment — the coffee cup and objects floating around the fall — was the hardest part to land. Camera movement and scene timing had to be written explicitly in the prompt, not implied.

I changed one variable per attempt so I could tell what actually improved the result. Changing three things at once felt productive and taught me nothing.

Sound and editing stayed in one place

Everything happened inside Higgsfield with Seedance 2.5 — the video, the sound direction, and the pacing. No separate sound-effects app and no external editor.

That kept the workflow short enough to repeat, which matters more to me than squeezing out the last few percent of polish. A workflow you can't repeat is a one-off, not a process.

What I'd change next time

More reference images for the location, and a shorter prompt. Long prompts feel thorough, but the model follows clear structure better than length.

I'd also keep a versioned log of each attempt from the start — not just the final prompt — so the reasoning behind each change survives instead of living in my head.

Keep reading

All monthly posts