Make Viral AI Short Films With This Exact Workflow — Full Tutorial

For the full breakdown, check out the YouTube video: I Made a Viral AI Short Film From Scratch — Full Workflow (Seedance 4K + Claude Fable 5). Every clip below is pulled straight from the tutorial, complete with the exact prompts!
Before You Start
Before generating a single video, build your assets — the character sheets, locations, and props. Every one of them will appear in multiple scenes, and a fixed reference is what keeps your hero's face, your props, and your locations identical from shot to shot.
Below we walk through the four core assets — the hero, the location, and the ships. The rest of the assets you'll get scene by scene.
Stage 1 — The Assets
The Hero — Pirate character sheet
What it does: Generates a 3-view character sheet of the hero as a pirate — full body front, full body back, and a frontal close-up — from your own reference photo. The close-up gives the model the exact face to lock onto; a grey background keeps the generations clean.
One important edit after that. A character sheet with multiple faces confuses Seedance later — it doesn't know which face to lock onto, and your hero slowly drifts scene by scene. Keep only the close-up face on the sheet:
The Location — Captain's cabin
What it does: Generates the pirate captain's cabin below deck — table, hanging lantern, map on the wall, thin rays of daylight leaking through the planks. Asking for a 3/4 angle adds depth, which gives the camera much more to hold onto when it moves.
The Props — The ships
What it does: Generates both ships on one split-screen sheet — front 3/4 and rear 3/4 of the same vessel — so the model always knows what the ship looks like from either side.
The bigger result became the enemy ship — but it needed to look dangerous. So we fed that bigger ship back into GPT Image 2 as the input image and ran an edit on top of it: open the sails, turn them black, add the skull emblem:
Stage 2 — The Prompting Framework
We don't write Seedance prompts by hand — a Claude Skill does it. Describe the scene, it splits it into shots and writes the full prompt, director hacks included.
Setup:
In Claude, go to Customize → Skills, hit + button, and upload the skill file.
Open a new chat, attach the script and every asset sheet, and give Claude the element list below.
In TAVOA AI, add each asset under Elements with exactly the same name.
This matters because during scene generation the prompts auto-match your inputs by these tags — nothing gets mixed up.
“@eduardo — main pirate character@loc_cabin — captain's cabin@main_ship_sheet — hero's ship@villain_ship_sheet — enemy ship@spyglass_sheet — spyglass@parrot_sheet — parrot@ship_cannon — ship cannon@map_prop — treasure map@cannonball — cannonball@button — the treasure case with the glowing green button@eduardo_crew — hero's crew@ocean_location — ocean@island_location — island@villain_captain — villain captain@helmsman_sheet — helmsman@main_character_desert — the hero in desert wanderer wardrobe@location_desert — desert dunes@rider — desert rider in black robes@ostrich — riding ostrich with goggles and saddle@binoculars — field binoculars@oasis — oasis with palms and a pond@main_character_jungle — the hero in explorer wardrobe@main_character_jungle_torn — the same explorer with claw-torn pants@location_jungle1 — jungle clearing with the colossal tree@location_jungle2 — ruined jungle path@location_jungle3 — jungle ridge over the cliff@spider — spider@macaque — white macaque@jaguar — jaguar@main_character_room — the hero in apartment wardrobe@cindy — Cindy, the little sister@room1 — apartment living room, master angle@room2 — the same room, reverse angle@room3 — the same room, desk wall angle”
From here, the workflow per scene is: describe what happens in plain language, attach the assets the shot needs, get the prompt, run it in Seedance 2.0, then take your notes back to Claude for exact fixes.
Stage 3 — The Scenes
Scene 1, Part 1
Here are all the assets used in this scene:
Assets built, skill installed — now the actual filmmaking. Scene by scene, here is every prompt that made the final cut. We open in the captain's cabin.What it does: Animates Eduardo searching the cabin, finding the treasure map, and reacting when a crewman bursts in with daylight behind him. One prompt covers the whole beat — the cuts, the dialogue line, and the sound come out of a single generation.
Every prompt from the skill follows one locked template: scene context → references → shot-by-shot action → lighting → locks. Audio is always real-world sound only — music gets added later, in the edit.
Scene 1, Part 2
What it does: Animates Eduardo running to the helm, taking the spyglass, and spotting the case on the island through a monocular crash-zoom POV — then the reveal of the enemy ship behind them. Giving the parrot a voice that screams the bad news twice is what makes the scene land.
Scene 1, Part 3
What it does: Animates the villain's introduction: the camera rises up the black-sailed hull to the captain, who delivers his line and fires the cannon himself in one unbroken take — then a hard cut to the miss beside Eduardo's ship. Merging the line and the shot into a single continuous take is what makes the sequence flow.
Scene 1, Part 4
What it does: Animates the near-miss barrage: water columns raining over the helm, the camera flying with the fired cannonball, and the slow-motion hits that hurl Eduardo toward the lens. When words fail on geometry, draw it — a quick sketch of where the shots land and where everyone stands tells the model more than ten sentences, and this prompt was built from exactly that drawing.
Scene 1, Part 5
What it does: Animates one continuous take: Eduardo flies into frame from the blast, fights down the burning wet deck through explosions, leaps over the bow — and the whole ship goes up behind him, the dust swallowing everything. Ending the shot inside the dust is what sets up the seamless transition into the desert.
Scene 2, Part 1
Here are all the assets used in this scene:
What it does: Animates Eduardo running to the helm, taking the spyglass, and spotting the case on the island through a monocular crash-zoom POV — then the reveal of the enemy ship behind them. Giving the parrot a voice that screams the bad news twice is what makes the scene land.
Scene 2, Part 2
What it does: Animates the sighting-and-flight beat: the hero spots the case through a binocular POV, launches down the dune face toward the oasis, and the riders close from a kilometer out with the storm wall growing behind them. One direction axis rules the whole prompt so he always runs toward the oasis. The oasis itself is a separate location sheet — our desert asset didn't have one, so we built it as its own asset, which keeps it identical in every generation.
Scene 2, Part 3
What it does: Animates the finale of the desert: the full chase to the oasis edge, the dive for the case — and the storm wall overrunning him mid-flight, meters short. The outcome lock is what sells the beat: the storm reaches the man before the man reaches the case, and almost is what makes it hurt.
Scene 3, Part 1
Here are all the assets used in this scene:
What it does: Animates a first-person wake-up — the frame opens from black through eye-slit shutters, the vertically flipped jungle blurred behind, and focus lands on the spider. Keeping the camera fully static and the spider off-center (fixed distance, at the frame edge) is what makes the POV read as an eye instead of a lens.
Scene 3, Part 2
What it does: Animates the fall-and-stampede gag: he trips, covers his head, begs mid-monologue — and the macaque troop simply floods over and past him. Timing the stampede to hit while he's still falling, plus a crane-down opening and a ground-level side profile for the chaos, is what makes the comedy land. The macaques here replaced a mandrill sheet that kept rendering cartoonish — swapping an asset mid-iteration is normal when a design fights the model.
Scene 3, Part 3
What it does: Animates one unbroken take: the jog along the ridge, the jaguar strike that swaps the hero to the torn wardrobe state, the chase to the cliff tree, and the fall into the gorge with the device locked in his fist while the jaguar stays at the edge. The torn wardrobe is its own second character sheet — whenever something on your character changes mid-story, generate the changed state as a separate sheet and it stops repairing itself between frames.
Scene 4, Part 1
Here are all the assets used in this scene:
What it does: Animates the twist reveal: the hero crashes from above onto the couch (stitching with the cliff fall), a giggle sounds behind him, and the swivel chair rotates to reveal Cindy with the device. The reverse-angle room references, the locked device size, and the parrot pinned to the top of the cage are what hold the shot-to-shot consistency.
Scene 4, Part 2
What it does: Generates the full shot-reverse-shot dialogue where Cindy counters the offer, presses the button, and the hero dissolves completely into green light. The camera is locked into two repeating framings — one identical clean single on Cindy for her lines, one identical over-the-shoulder toward the hero for his — so the 180° axis holds across every cut and the scene assembles cleanly in the edit.
Want to Make Viral AI Short Films?
by Mariam Barova