Key takeaway — A good video prompt isn't a long prompt; it's a sorted one. ⑥ Pick one of three shapes: section blocks, a timecoded shot list, or one continuous take ⑦ Make the timecodes add up exactly to the length ⑧ One action per shot ⑨ Nouns and numbers instead of adjectives ⑩ Keep block names stable. Of 60 publicly shared AI video prompts, 54 (90%) used one of the three shapes.
This is part 3 of our AI ad video directing series. If part 2 settled length, purpose, ratio, budget, and references, it's time to decide what shape the prompt itself should take. This is the second group of the xbrush Academy's 20 Tips.
Tip 6. Pick one of three shapes
Section blocks, a timecoded shot list, or one continuous take. Write freely and the model wavers over which lines are instructions and which are atmosphere. Commit to a shape and the reading is fixed — and when you revise, you know which line to touch.
Shape | When to use it | Of 60 |
|---|---|---|
Section blocks | To hold mood, color, camera, and people separately | 24 |
Timecoded shot list | Multiple cuts — one line per mark | 21 |
One continuous take | Running unbroken — written in order of action | 9 |
Real example — section blocks
Name each block in brackets and fill in the values beneath it. These are the blocks of a medieval street procession prompt:
[STYLE + CAMERA + ATMOSPHERE]
[CHARACTERS]
[LOCATION]
[TIMELINE]
[STYLE & QUALITY BOOSTERS]Source: X @techhalla (x.com)
Real example — timecoded shot list
A luxury pop-duo music video splits 30 seconds into six five-second scenes, each with one block of location and action.
Scene 1 (0–5s)
Inside a luxurious pastel studio featuring giant LED walls, glossy reflective floors ...
Scene 2 (5–10s)
Golden-hour rooftop overlooking a modern skyline ...Source: X @ZaraIrahh (x.com)
Real example — one continuous take
For an uncut video, declare "no cut" in the first line and then write only in order of action.
SEQUENCE SHOT. NO CUT.
0-4s: She rises from the sofa empty-handed, walks toward a kitchen table lined with
alcohol bottles and glasses. Handheld camera follows close behind her shoulder, slightly shaky.Source: shared by X @matthieu_ai (x.com)
Ads usually mix section blocks with timecodes: blocks at the top lock person, product, light, and color, and a TIMELINE block at the bottom splits the shots by time.
Tip 7. Make the timecodes add up to the length
Overshoot the total and the model trims the back end. Shot lists creep. When the segments add up past the stated length, the model quietly shortens something — usually the last shot, which is exactly where the end card or the product close-up lives.
Add the segments up once when you're done.
Over? Drop a shot rather than shaving every shot a little.
Leave about half a second of slack on the final shot so the ending doesn't clip.
Weaker | Stronger |
|---|---|
0:00–0:05 / 0:05–0:12 / 0:12–0:20 (a 15-second film totalling 20) | 0:00–0:04 / 0:04–0:10 / 0:10–0:15 (totals 15) |
Real example — a timeline joined to the decimal
The noir scene in the Higgsfield guide makes the end of each segment meet the start of the next exactly, down to the decimal.
ACTION TIMING: 0.0s to 6.0s, SEGMENT 1, 84° wide. ... HARD CUT.
6.0s to 12.0s, SEGMENT 2, 29° medium. ... HARD CUT.
12.0s to 19.0s, SEGMENT 3, 47° normal. ...Source: Higgsfield Blog — Seedance 2.5 Prompting Guide (higgsfield.ai)
Tip 8. One action per shot
Two things happening in three seconds means neither reads. Stack actions into a short shot and the model starts them all at once. Someone opening a door while turning and smiling gives you none of the three cleanly.
Two verbs in one line? Split the shot.
Count camera movement as an action — a walking subject plus a moving camera is two.
If they truly must overlap, state the order: "turns, then smiles."
Weaker | Stronger |
|---|---|
She opens the door, walks in, looks at camera and smiles while setting the cup down. | Shot 1: opens the door. Shot 2: sets the cup down. Shot 3: looks at camera. |
Real example — one named action per segment
Higgsfield's 20-second headphone ad gives every segment a single-action name — "the walk," "the find," "the touch" — and even notes the second each action happens.
Segment 1 (0.0s to 4.0s), the walk, she strolls the lawn path in the sun ...
Segment 2 (4.0s to 8.0s), the find, over-shoulder push-in as she approaches the floating headphones ...
Segment 3 (8.0s to 10.5s), the touch, tight on her fingertips meeting the shell at 8.6s ...Source: Higgsfield Blog (higgsfield.ai)
The fantasy action example in the same guide cuts 30 seconds into 17 shots. The common pattern: the faster the action, the more shots — and the less each shot has to do.
Tip 9. Don't push with length
A good prompt isn't long. It's sorted. Piled-up adjectives fight each other. "Warm and calm and bold and minimal" scatters the direction four ways. The same word count reads far more reliably when every value sits in a known slot.
Fewer adjectives, more nouns and numbers.
One value per field — color in the color slot, camera in the camera slot.
Read it back; if two lines contradict, delete one.
Real example — 30 seconds in about 1,600 characters
The median length across the 60 prompts was about 4,000 characters, but some ran past 16,000 and others came in around 1,600. The New York bodega example in fal.ai's prompting guide covers a 30-second continuous take in roughly 1,600 characters. What it has instead is who, which hand, and what in every sentence.
The same bike messenger wears a yellow rain jacket and carries one red bicycle helmet
in the left hand throughout. 0-5 seconds: the door bell rings as the messenger enters,
closes the glass door with the right hand, ...Source: fal.ai Learn — Seedance 2.5 Prompting Guide (fal.ai)
Length varied widely from strand to strand. What held constant wasn't length — it was that the fields were separated.
Tip 10. Keep your block names stable
Reuse the same names and you always know where to edit. Block names are signposts for the model, but mostly they're edit points for you. Use the same names on the next film and last film's values carry straight over — and you can see what you changed and what it did.
Settle on five or six names — e.g., Subject / Action / Space / Camera / Light / Sound.
Keep the slot even when it doesn't apply this time — write "none."
Keep the versions that worked and start the next film from them.
Real example — a fixed seven-field form
The open-source Veo 3 prompting guide on GitHub writes every example with the same seven fields. Even a scene without dialogue keeps the slot.
Subject: ...
Action: ...
Scene: ...
Style: A medium shot captures RedShirtGuy at eye-level, on a static tripod, ... 16:9 aspect ratio.
Dialogue: (Authoritative, instructional) ...
Sounds: ... the faint squeak of a marker on the whiteboard, and a quiet hum of air conditioning.
Technical (Negative Prompt): subtitles, captions, watermark, text, ...Source: GitHub — snubroot/Veo-3-Prompting-Guide (github.com)
What's interesting is that block names spread beyond their authors. Among the 60, nine prompts opened with SCENE CONTEXT and eight used a POSITIVE LOCKS block — not only in Higgsfield's official examples but in prompts individual creators posted on YouTube, Notion, and Threads. People adopt a form that worked, names and all, and change only the values.
Try it on xbrush — a six-slot template
Save this template and change only the values each time you hand it to the xbrush workspace agent or Cinema. It's a hybrid: section blocks (top five slots) plus a timecoded action slot.
[Subject] The tumbler in attached image 1 — keep shape, color, and logo
[Space] Kitchen counter in the morning, white tile wall
[Camera] Close-up → medium shot, eye level, locked off
[Light] 9am, soft side light from the window on the right
[Sound] Diegetic: lid click (0:05). Music: none
[Action] 15 seconds, 4 shots
0:00–0:04 A hand picks up the tumbler
0:04–0:08 Thumb flips open the one-touch lid
0:08–0:12 Takes a sip
0:12–0:15 Tumbler front on, on the counter, locked off (0.5s slack at the end)Next up: how to lock what goes inside those slots — people, camera, light, and color.
Frequently Asked Questions
What structure should an AI video prompt follow?
Pick one of three shapes: section blocks, a timecoded shot list, or one continuous take. Of 60 publicly shared prompts, 90% used one of these. Multi-shot videos like ads often use a hybrid, with blocks for people, light, and color at the top and timecoded shots below.
Do I have to write timecodes?
With two or more shots, writing them is more reliable. When you do, the segments must add up exactly to the total length. If they overshoot, the model trims something on its own — usually the last shot, where the end card or product close-up sits.
Are longer prompts better?
No. The median length of 60 published prompts was about 4,000 characters, but there are good 30-second prompts of around 1,600. What matters more than length is that fields are separated and direction is given in nouns and numbers rather than adjectives.
Why shouldn't I put several actions in one shot?
When actions are stacked in a short shot, the model starts them all at once and none of them reads cleanly. If a line has two or more verbs, split the shot — and count camera movement as an action too.
Do block names have to be in English?
No. What matters is using the same names consistently. That said, most public examples use English block names like SCENE CONTEXT and POSITIVE LOCKS, so English names can be convenient when you borrow from or adapt other people's prompts.