Two characters. One of them is me. Neither of us was ever filmed.
A 100-second film with dialogue and directed cameras — planned in 3D first, generated in Seedance 2.5 after. This is the whole path, step by step.
Start with an idea you can say in one sentence.
Mine was: it's not the club.
A guy on a golf course blames the club, then the ball, then his luck — everything except the fact that he walked up to the tee with no plan. Swap golf for AI video and you get everyone who blames the model, the prompt and bad luck.
Before I opened a single tool, the film existed as words: two men, one tee, one page of dialogue. Every line already had a job. If a scene doesn't work on the page, no model will fix it — it will only make the problem look expensive.
GUY Nice shot.
ANDREIWanna try?
GUYAfraid that's beyond me. Especially without gear like that.
ANDREIIt's not the club. It's seeing the shot before you take it.
GUY"Seeing the shot." My brother-in-law says that about crypto. He lives in a studio apartment now.
So the first file of this film is not a prompt. It's the script.
Build the place before the film
Most AI video starts with a prompt and a hope. This one started with a grey box.
I built the whole location once, in 3D, inside Higgsfield 3D Jutsu: the tee, the green, the dunes, the clubhouse, the ocean. Two mannequins. Beige is me, blue is Guy. No textures, no faces. Just space.
Then every shot became a camera inside that one scene. Same tee, same sun, same distance to the flag from every angle. That's why the cuts match: no model had to guess where the clubhouse is, because it never moved.
The mannequins walk, turn and swing on a timeline. 92 seconds of grey animation.
I didn't model any of it by hand. Claude wrote the scene in code. I moved cameras and said "lower", "closer", "wrong side".
Watch it grey. Fix it grey.
A grey render cost me zero credits. A wrong final costs credits and an evening.
So I watched the grey version like a real film and fixed it there:
- Guy walked in flat from the side. Now he comes up from the bottom corner on a diagonal.
- We stood on the wrong sides, and every angle needed an awkward turn. Swapped.
- Guy started moving three seconds late. Now he's already walking while the camera is still on the high shot.
- Nobody could see the sheet of paper the whole joke rests on. One new close-up camera, 50 mm.
The scene ended up with 33 shots on one timeline. By the time it reached the video model, every camera, step and turn had a timecode.
Give the model a job it can't misread
The final is Seedance 2.5 on Higgsfield. The scene is cut into 8 clips, 10 to 14.5 seconds each. Every generation gets four files:
- the grey video of that clip: camera, framing, timing
- my character sheet
- Guy's character sheet
- one location plate
And a prompt in four blocks: MATERIALS, CONCEPT LINE, TIMELINE, GLOBAL LOCK.
The timeline is not "he swings, then a guy walks up". It's cuts at 1.5, 2.5, 3.0, 4.5 and 8.5 seconds, impact at 2.6, Guy stops at 8.1. The numbers come straight from the 3D scene.
Here is that block for the opening clip, in full:
③ TIMELINE
0-1.5s. Image: ground-level close shot, exactly as @video1: the white ball on the tee in the foreground, ANDREI towering above it, seen from below, the driver shaft making a diagonal. Camera: the lens and framing of @video1, almost static, a tiny drift as in @video1. Action: ANDREI settles over the ball, calm and focused; at about 1 s he starts taking the driver back. Sound: sea wind across the dunes.
1.5-2.5s. Hard cut. Image: low three-quarter shot, exactly as @video1: the golf bag big in the foreground on the left, ANDREI at the top of his backswing. Camera: the lens and framing of @video1, a slight push-in. Action: ANDREI swings the driver back to the top and starts the downswing. Sound: the swish of the backswing.
2.5-3.0s. Hard cut. Image: extreme close-up on the ball on the tee, exactly as @video1. Camera: locked-off. Action: at about 2.6 s the driver head smashes into the ball and the ball shoots out of frame. Sound: a sharp powerful crack of the driver on the ball.
3.0-4.5s. Hard cut. Image: high wide shot from behind the tee straight down the hole, exactly as @video1: ANDREI small in the lower part of the frame finishing his follow-through, the fairway, dunes, bunker and green below, the ocean along the upper third, the clubhouse on the right. Camera: the lens and position of @video1, a slow push-in. Action: ANDREI's follow-through ends at about 3.3 s and he holds his finish; the ball flies away toward the distant green; we do not see where it lands. Sound: a fading whistle of the ball, sea wind.
4.5-8.5s. Hard cut. Image: wide shot from the side of the tee, exactly as @video1: ANDREI on the left third holding his finish, the ocean behind; from about 4.6 s GUY walks into the frame from the bottom right on a diagonal, unhurried, and at about 8.1 s stops on the right third, about three metres from ANDREI. Camera: the lens and position of @video1, a slow steady push-in. Action: still walking, GUY calls out, impressed and a little cocky, with a light rising tone: "Nice shot." At about 6 s ANDREI turns his head toward GUY, then from 6.2 s to 7.3 s turns his body to him, and says, dry and unhurried, a hint of a smile: "Wanna try?" GUY stops and turns to face ANDREI. Sound: footsteps on short grass, sea wind.
8.5-12.0s. Hard cut. Image: over ANDREI's right shoulder, exactly as @video1: the back of ANDREI's head and shoulder as a dark mass on the left, GUY facing him on the right third. Camera: the lens and framing of @video1, static, then at about 10.7 s, on the words "gear like that", one fast zoom-in onto GUY and stop. Action: GUY, grinning, relaxed and teasing, says: "Afraid that's beyond me." He waves a hand at ANDREI's driver and adds, amused, eyebrows raised: "Especially without gear like that." Silence for the last half second. Sound: sea wind gusting.
When a clip came out 90% right, I didn't re-roll it. Four 5-second fixes patched the bad moments.
This is how the prompt for the opening clip begins:
① MATERIALS
@video1 gives the camera, the framing, the layout and the timing: the six camera set-ups and the cuts between them, camera position and lens in each, every camera move, where each man stands and walks, who is left and who is right, when ANDREI swings and strikes the ball, when GUY walks in and when ANDREI turns to him. The grey and coloured shapes in @video1 are spatial guides only: the beige figure becomes ANDREI, the blue figure becomes GUY, and every shape is replaced completely by real people, a real tee, real golf clubs, real grass, real dunes and real buildings; nothing in the video looks like a 3D model.
The place is the links course shown in @image3 — use it for the look of the land, the green, the…
The rest of the prompt, and the whole kit behind the film:
The prompts $5.99
The two full scene prompts from this film: one driven by the grey video, one from a first frame. Plus the character sheet and location prompts, and an adapter that makes Claude or ChatGPT rewrite them for your scene.
Get the promptsThe full guide $29
The prompts show what the model was told. The guide shows how to get there with your own film:
- the request I send Claude to make a scene land
- a storyboard prompt to see the film before you build it
- the brief template for the 3D Jutsu agent, and mine in full
- what to fix in grey, for 0 credits, before a single generation
- getting the driver videos out, and the upload order for Seedance 2.5
- one clip taken apart, grey to final, and how I fixed what broke
A 15-page PDF + the whole prompt pack.
Get the full guideWhat went wrong on the way
Not everything worked the first time. These are the real failures, and what each one taught me.
- The dialogue took eleven versions. The early ones were flat: lines that only moved the plot along. Every rewrite happened on the page, where it cost nothing.
- A golf club turned into a black cube. When Claude copied a club inside the 3D scene, it lost its scale and rendered as a giant black block. In grey, a one-minute fix. In a final, a wasted generation.
- The clubhouse came out wrong. I hadn't given the model the location plate, so it invented its own building. A model can't match what you don't show it. Since then every generation gets the plate.
- One cut gave the trick away. After the snap, the next shot came from a different clip: I was a different size, and we stood one metre apart in one shot and three in the next. I regenerated that whole stretch, from "I wrote it" to the strike, as one 13-second clip with one wide camera.
What do I need to repeat this?
A Higgsfield account for 3D Jutsu and Seedance 2.5, and Claude or ChatGPT. That's the whole stack this film was made on.
Do I need to know 3D?
No. I didn't model anything by hand. Claude wrote the scene in code, I moved the cameras.
My video isn't about golf. Will this work?
The prompts are the real ones from this film, word for word. The adapter tells Claude or ChatGPT how to rewrite them for your characters, your place and your timing. The structure stays, the golf goes.
What if I don't want to build a 3D scene?
The second prompt in the pack works without one. It starts from a single first frame.
Which one should I get?
The prompts show you what the model was told. The full guide shows how to get there with your own film: from the idea and the 3D scene to the grey check and the final clips, with every mistake. And it includes the prompts.