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

2026-07-14 · blog how-to · Higgsfield · original

Topic: Video generation & shot direction (also: Assets, references & consistency, Prompting & AI assistants, Audio, voice & music) · mode: Tutorial · tags: prompts shared, one creator's experience

Higgsfield's blog breaks down how it made a viral short film, starting from the rule that character, location and prop sheets come before any video so every shot matches. Soul Cinema gives the raw pass, GPT Image 2 the clean sheet, and a Claude Skill writes every Seedance 2.0 prompt in one template, with assets named exactly as in Claude's element list. Scene by scene it adds fixes: one face per character sheet, a new sheet when a costume changes, sketches for geometry, and two locked framings for dialogue.

What is said

Want to Make Viral AI Short Films?

What is said

★ — how much it matters to the source.

★★ Higgsfield's blog breaks down the short film from its YouTube tutorial, with every clip taken from the tutorial and the exact prompts.
★★★ Before generating any video, Higgsfield builds the film's assets (character sheets, locations and props), since each appears in several scenes and a fixed reference keeps faces, props and locations identical from shot to shot. · character sheet
★★★ Higgsfield's asset routine: Soul Cinema for the raw pass (simple prompt, lots of variety), GPT Image 2 for edits and the final clean sheet; then the Claude Skill writes every scene and Seedance 2.0 generates it. · Soul Cinema, GPT Image 2, Seedance 2.0, Claude
★★ The hero's pirate character sheet has three views made from your own reference photo: full body front, full body back and a frontal close-up that gives the model the exact face, on a grey background that keeps generations clean. · character sheet
★★★ A character sheet with several faces confuses Seedance 2.0, which does not know which face to lock onto, so the hero drifts scene by scene; Higgsfield keeps only the close-up face on the sheet. · character sheet, Seedance 2.0
★★ For the captain's cabin location, Higgsfield asks for a 3/4 angle, which adds depth and gives the camera much more to hold onto when it moves. · location reference sheet
★★ Higgsfield generates the 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.
★★ To make the enemy ship look dangerous, Higgsfield fed the bigger ship back into GPT Image 2 as the input image and edited it: open the sails, turn them black, add the skull emblem. · GPT Image 2
★★★ Higgsfield does not write Seedance 2.0 prompts by hand: you describe the scene to a Claude Skill, which splits it into shots and writes the full prompt, director hacks included. · Claude, Seedance 2.0
★★ To install the skill, in Claude go to Customize → Skills, press the + button and upload the skill file; then open a new chat with the script, every asset sheet and the element list attached. · Claude
★★★ In Higgsfield, add each asset under Elements with exactly the same name it has in the element list given to Claude. · Elements, Claude
★★ During scene generation the Claude Skill's prompts match the inputs by these Elements tags, so nothing gets mixed up. · Elements, Claude
★★ The film's element list gives each asset an @ tag with a short description, such as "@main ship sheet — hero's ship" and "@villain ship sheet — enemy ship". · Elements, Claude
★★★ Higgsfield's per-scene loop: 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. · Claude, Seedance 2.0
★★ In the film's opening in the captain's cabin, one Seedance 2.0 prompt covers the whole beat: the cuts, the dialogue line and the sound come out of a single generation. · Seedance 2.0
★★★ Every prompt from Higgsfield's Claude Skill follows one locked template: scene context, references, shot-by-shot action, lighting, then locks. · Claude
★★★ In Higgsfield's film prompts the generated audio is always real-world sound only; music is added later, in the edit. · native audio
★★ For the villain's introduction, Higgsfield merges his line and the shot (camera rising up the hull to the captain, who speaks and fires the cannon) into one continuous take, which it says makes the sequence flow.
★★★ When words fail on geometry, Higgsfield draws it: a quick sketch of where the shots land and where everyone stands tells the model more than sentences, and the near-miss barrage prompt was built from such a drawing.
★★★ Higgsfield ends the ship-explosion take inside the dust, which sets up a seamless transition into the desert scene.
★★ In the desert chase prompt one direction axis rules the whole prompt, so the hero always runs toward the oasis.
★★★ The desert asset had no oasis, so Higgsfield built the oasis as its own location sheet, which keeps it identical in every generation. · location reference sheet
★★ An outcome lock sells the desert finale: the prompt fixes that the storm reaches the man before the man reaches the case.
★★ For a first-person wake-up shot, Higgsfield keeps the camera fully static and the spider off-centre at a fixed distance at the frame edge, so the POV reads as an eye instead of a lens.
★★ For the jungle stampede gag, timing the macaques to hit while the hero is still falling, with a crane-down opening and a ground-level side profile, is what Higgsfield says makes the comedy land.
★★★ Higgsfield replaced a mandrill sheet that kept rendering cartoonish with macaques, saying swapping an asset mid-iteration is normal when a design fights the model.
★★★ When something on a character changes mid-story, such as a claw-torn wardrobe, Higgsfield generates the changed state as a separate character sheet, and it stops repairing itself between frames. · character sheet
★★ In the apartment twist, reverse-angle room references, a locked device size and the parrot pinned to the top of its cage hold the shot-to-shot consistency.
★★★ For shot-reverse-shot dialogue, Higgsfield locks the camera into two repeating framings, an identical clean single on Cindy and an identical over-the-shoulder toward the hero, so the 180° axis holds across every cut and the scene assembles cleanly.

Threads in this video

refine · How do you keep a character the same across shots? — A character sheet with several faces confuses Seedance 2.0, which does not know which face to lock onto, so the hero drifts scene by scene; Higgsfield keeps only the close-up face on the sheet.
restate · How do you keep a character the same across shots? — When something on a character changes mid-story, such as a claw-torn wardrobe, Higgsfield generates the changed state as a separate character sheet, and it stops repairing itself between frames.
refine · How do you keep a location, product or prop the same across shots? — The desert asset had no oasis, so Higgsfield built the oasis as its own location sheet, which keeps it identical in every generation.
restate · How do you keep a location, product or prop the same across shots? — Higgsfield generates the 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.
refine · Which image model should be used for keyframes, edits, text and outfits? — Higgsfield's asset routine: Soul Cinema for the raw pass (simple prompt, lots of variety), GPT Image 2 for edits and the final clean sheet; then the Claude Skill writes every scene and Seedance 2.0 generates it.
restate · Which LLM should be used to write video prompts? — Higgsfield does not write Seedance 2.0 prompts by hand: you describe the scene to a Claude Skill, which splits it into shots and writes the full prompt, director hacks included.
restate · When a generation fails, should you generate more takes, fix the prompt or start over? — Higgsfield's per-scene loop: 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.
refine · When a generation fails, should you generate more takes, fix the prompt or start over? — Higgsfield replaced a mandrill sheet that kept rendering cartoonish with macaques, saying swapping an asset mid-iteration is normal when a design fights the model.
refine · How do you direct a multi-shot sequence? — Every prompt from Higgsfield's Claude Skill follows one locked template: scene context, references, shot-by-shot action, lighting, then locks.
restate · How do you direct a multi-shot sequence? — In the film's opening in the captain's cabin, one Seedance 2.0 prompt covers the whole beat: the cuts, the dialogue line and the sound come out of a single generation.
restate · Should sound and music be generated with the clip or added in the edit? — In Higgsfield's film prompts the generated audio is always real-world sound only; music is added later, in the edit.
refine · How do you get precise motion (dance, action, choreography)? — When words fail on geometry, Higgsfield draws it: a quick sketch of where the shots land and where everyone stands tells the model more than sentences, and the near-miss barrage prompt was built from such a drawing.
refine · How do you make a transition between two shots? — Higgsfield ends the ship-explosion take inside the dust, which sets up a seamless transition into the desert scene.
refine · How do you control where the camera is in a known set? — For shot-reverse-shot dialogue, Higgsfield locks the camera into two repeating framings, an identical clean single on Cindy and an identical over-the-shoulder toward the hero, so the 180° axis holds across every cut and the scene assembles cleanly.

Mentioned

Central: Claude · Claude Skill for Seedance 2.0 scene prompts (viral short film) · Higgsfield's viral AI short film with Eduardo (2026) · Seedance 2.0

Also: Concepts: character sheet · location reference sheet · native audio · product sheet

Characters: Cindy · Eduardo

Products: Claude Fable 5 · Elements · GPT Image 2 · Soul Cinema

Also centred on Claude Skill for Seedance 2.0 scene prompts (viral short film): I Made a Viral AI Short Film From Scratch — Full Workflow (Seedance 4K + Claude Fable 5) · Blockbuster 4K: The AI Filmmaking Pipeline: Stage 3: Scene Generation with the Prompt-Builder Skill

Also centred on Higgsfield's viral AI short film with Eduardo (2026): I Made a Viral AI Short Film From Scratch — Full Workflow (Seedance 4K + Claude Fable 5) · Blockbuster 4K: The AI Filmmaking Pipeline: Full Film and Workflow Recap · Blockbuster 4K: The AI Filmmaking Pipeline: One Character, Many Worlds: Watch the Film · Blockbuster 4K: The AI Filmmaking Pipeline: Stage 3: Scene Generation with the Prompt-Builder Skill · Blockbuster 4K: The AI Filmmaking Pipeline: Stage 2: Building Character, Location, and Prop Assets · Blockbuster 4K: The AI Filmmaking Pipeline: Stage 1: Writing the Script with Claude

Replies to

refers to · I Made a Viral AI Short Film From Scratch — Full Workflow (Seedance 4K + Claude Fable 5)

Replied to by

← Previous: 7 Best All-in-One AI Tools for Video, Audio & Editing in 2026 · Next → How One of One, the Studio Behind the Official Maradona Label, Took a Brand World From Concept to Sold Out With Higgsfield