39:28
AI filmmaking workflow
The video presents a full pre-production-to-post workflow for building a coherent, cinematic AI short film in Seedance (Higgsfield's 4K model) combined with a custom Claude prompt-writing skill, arguing that meticulous asset preparation — named/tagged character and location sheets, consistency techniques like reverse-angle references and the 180° rule — is what separates a cinematic, credit-efficient result from cheap-looking one-off generations.
The demo film spans three very different settings (storm at sea, desert, jungle) tied together by one story with a twist ending, all generated in Seedance at 4K.
The script wasn't produced by one request; Claude was told to expand a one-line idea and ask clarifying questions (runtime, settings, plot twists) before drafting.
Character sheets combining front view, side profiles, close-up face, and full body on a gray background keep the model locked onto one consistent face.
Multiple faces in a single reference image cause facial drift by around scene 5; erasing the face from the full-body pose leaves only the close-up as the identity anchor.
Locations are treated as make-or-break: a cheap-looking location makes every shot inside it look cheap, and no later prompting fixes that.
3/4-angle location references outperform flat head-on shots for win rate and give the camera more to move around.
Sub-locations (e.g., an oasis inside the desert) are generated and stored as their own separate assets so they render identically whenever they reappear.
Every asset is named, introduced to Claude by that name, and uploaded to Higgsfield as an 'element' with the matching name so scene prompts automatically match the right inputs.
A custom Claude skill splits a scene into shots and writes detailed Seedance prompts (21x9, 4K, 15 seconds) from the script and assets, rather than the creator hand-writing prompts.
When words fail to convey blocking or camera geometry, a simple drawing of positions communicates it better than a text description.
Small character reactions (a smile, slamming a map down, moving toward a door) are what make a shot feel real.
Batches are run four at a time; failures are treated as data, and the best parts of different batches are combined into the final cut.
Continuity breaks (cartoonish monkeys, self-repairing torn pants, shifting room backgrounds) are fixed with new reference assets, not just reworded prompts.
Dialogue scenes require reverse-angle environment references of the room to stop backgrounds, prop sizes, and object placement from jumping between cuts.
Camera continuity in dialogue is locked with the 180° rule: the camera stays on one static side, and each character is filmed from a consistent angle, or the audience loses track of who is standing where.
The three hardest/most important shots get the most iteration; once solved, the same established approach carries the remaining, easier shots.
The full deliverable (script, every prompt, asset sheets, the Claude skill) is shared publicly, and the presenter frames the workflow as portable to commercials or any other short-film idea.
Idea-expansion scripting with Claude — A scripting method where instead of asking Claude to 'write me a script,' the creator gives it a bare idea and lets it ask clarifying questions about runtime, settings, and plot twists before drafting. Apply: Feed Claude a one-line premise and answer its follow-up questions rather than demanding a finished script in one request.
Character sheet (multi-angle reference) — A single reference image combining a character's front view, left/right side profiles, close-up face, and full body front-and-back to lock the model onto one consistent face. Apply: Generate one full character sheet per character before generating any scene shots involving them.
Gray-background pro tip — Using a plain gray background behind character sheets removes background clutter and raises the generation 'win rate.'. Apply: Set character-sheet backgrounds to gray/plain when generating reference images.
Single-face erasure fix — A fix for facial drift caused by multiple faces in one character-sheet image (drift observed by scene 5); erasing the face from the full-body pose leaves only the close-up as the identity anchor. Apply: When a character sheet shows more than one visible face, erase the face from the full-body pose so only the close-up remains.
3/4 angle for locations — Requesting locations at a 3/4 angle rather than flat head-on gives more depth, a higher win rate, and something for a moving camera to hold on to. Apply: Prompt location references at a 3/4 angle instead of a straight-on view when building location assets.
Soul Cinema raw-pass / GPT Image 2.0 edit-pass split — A two-tool asset pipeline: Soul Cinema for a cheap raw pass with a simple prompt to get many varied options (about 8 images per credit), then GPT Image 2.0 for edits and final clean sheets. Apply: Generate a wide variety of raw options cheaply in Soul Cinema, pick the best, then clean and finalize it in GPT Image 2.0.
Named/tagged asset system — Naming every asset, introducing it to Claude by that name, and uploading it to Higgsfield as an 'element' with the exact same name so scene prompts auto-match the right inputs. Apply: Give every character/location/prop a consistent name and reuse that identical name across Claude conversations and Higgsfield element uploads.
Asset folder organization / Canvas workspace — Organizing generations into scene folders with subfolders per asset, and keeping finalized assets visible together in Higgsfield's Canvas, cited as what 'pays off the most' over hundreds of generations. Apply: Set up a scene-to-asset folder structure from the start and keep chosen final assets visible in one workspace for quick reference during shot generation.
Claude prompt-writing skill (for Seedance) — A custom Claude skill that splits a scene into shots and writes detailed Seedance prompts from the script and uploaded assets. Apply: Activate the skill, upload the shot's required assets plus the script, and let it generate the detailed shot prompt instead of hand-writing it.
Cinematic shot parameters (21x9 / 4K / 15s) — A standard technical configuration used for every shot: 21x9 cinematic aspect ratio, 4K resolution, and 15-second clip duration. Apply: Set these three parameters as the default for every Seedance generation to keep the film visually consistent.
Batch generation (run 4, pick/combine best) — Running four generations per shot as a default and either selecting the best one or combining the best parts across different batches. Apply: Generate at least four variations before judging a shot, and mix elements from different batches rather than expecting one perfect take.
Visual reference for blocking/geometry — Giving the model a simple drawing of character or camera positions instead of describing spatial geometry in words. Apply: When a shot needs precise blocking or a specific camera angle, sketch the positions and feed that image to the model alongside the text prompt.
Location-within-location asset isolation — Treating a distinct sub-location (e.g., an oasis inside the desert) as its own separate asset so it renders identically every time it reappears. Apply: Generate and save a dedicated asset image for any recurring sub-setting rather than folding it into the main location prompt.
Three-key-shots prioritization — Identifying the three most important/hardest shots that hold the story together, solving those first, then applying the same established logic to the remaining shots. Apply: Flag the handful of shots the story can't work without and iterate on those until solved before batch-producing the rest.
Continuous POV shot for failure recovery — Switching to a single continuous point-of-view take as a fallback after multi-shot attempts at a POV sequence kept misfiring (pushing objects into a literal eye, or the camera zooming out like a physical lens). Apply: If a model keeps literalizing 'POV' across a multi-shot sequence, collapse it into one continuous POV take instead.
Off-center jump-scare framing — Positioning a scare element off to the side of the frame rather than dead-center made the jump scare read as more natural. Apply: Frame a reveal/jump-scare object off-center rather than centered so it feels like a natural discovery.
Cuts-driven comedy — A stated principle that comedic effect in a multi-shot sequence depends on the cuts between shots, not just single-shot content. Apply: Break comedic beats into multiple distinct shots with cuts between them rather than one continuous take, to preserve comic timing.
Mid-production asset substitution — Swapping out an asset (e.g., replacing cartoonish monkey designs) mid-production when a generation looks too plastic or the action reads unclear, with backup variants kept ready. Apply: Keep backup variants of a problematic asset on hand and swap the asset itself, not just the prompt, then re-run batches with the new reference.
Character-state reference sheet — A fix for state-continuity errors (e.g., ripped pants healing themselves) by generating a second character sheet showing the character already in the changed state and feeding it in as an additional reference. Apply: When a character's appearance changes partway through the story, create and feed the model a dedicated reference sheet of that new state.
Reverse-angle environment references — For dialogue scenes cut between characters, generating images of both sides of the room gives the model a 'blueprint' so it stops changing the background, prop sizes, or object placement between cuts. Apply: Before shooting a dialogue exchange, generate reference images looking both directions across the set and feed both into every shot in that scene.
180° rule (camera-axis lock) — A camera-continuity rule keeping the camera on one static side of the action, filming each character from a consistent angle/shoulder so the audience keeps a stable sense of who is standing where. Apply: Instruct the model explicitly to keep the camera on one fixed side of the dialogue axis and to always frame the reverse shot over the same shoulder.
Prompt-building framework — The overall process of writing a full script first, then running it through the Claude skill to convert it into highly detailed per-shot prompts, rather than issuing one-line prompts. Apply: Produce a complete script before prompting, then feed it plus assets through a structured prompt-conversion step rather than prompting shot-by-shot from scratch.
The POV shot's repeated failures reveal a specific AI-video limitation: the model can literalize 'camera as eye' instructions by trying to push objects toward a rendered eyeball rather than treating the lens itself as the character's eye.
Off-center framing of a jump-scare element reads as more natural than dead-center framing, suggesting composition choices affect perceived realism independent of the prompt's content.
The self-healing continuity error (ripped pants re-stitching between shots) was fixed not with better wording but with a new reference image encoding the damaged state, implying persistent state changes need to live in the asset library rather than the prompt text.
Comedic timing is treated as a function of the cuts between shots rather than shot content alone, which is why a monkey-attack beat was deliberately broken into multiple shots instead of kept as one continuous take.
Background/prop drift across a dialogue scene's reverse cuts (a parrot appearing in vs. on top of its cage, a device changing size) is solved the same way facial drift is solved for characters — by feeding the model bidirectional reference images so a 'room' is treated as its own consistency asset.
Near-misses ('almost') are named as the deliberate dramatic engine of the escape sequence rather than the escape itself.
Two weeks and roughly 400 generations is presented as the normal cost of producing a ~15-second narrative short, with organization (naming, folders, Canvas) — not prompting cleverness — cited as what makes that volume tractable.
«Hi, I'm Adil and what you just watched is an action short film made in sea dance in 4K. Crazy, right?»
— 00:22
«Nothing good comes out of one lazy request. The details are what make it work. That's why we get a script that gets millions of views.»
— 04:45
«Locations carry the whole film. If the setting looks cheap, every shot inside it does too. And no amount of prompting fixes that later.»
— 07:31
«When the character actually reacts to what's happening around him, the whole shot starts to feel real.»
— 13:51
«One drawing tells the model what 10 sentences can't.»
— 17:29
«Almost is what makes it hurt.»
— 23:43
«Now that is a real POV. I love how his eyes slowly struggle to open. And keeping the spider off to the side makes the whole jump scare feel way more natural.»
— 26:47
«When making a dialogue scene with reverse angles, you must generate reverse angle environment references of the same room. Feed the model images of what both sides of the room look like. That gives the model a blueprint of the room and the background stops jumping between cuts.»
— 32:35
«So, that's the 180° rule. break it and the audience stops knowing who's standing where.»
— 34:18
«We approached this like actual filmmakers. We didn't just throw a random oneline prompt and hope for the best. We built a script first and then used our prompt building framework to turn it into highly detailed working shots.»
— 38:26
«The best part is that this exact workflow works for anything. Commercials, short films, or any random idea in your head. It's the exact same approach.»
— 39:01
Reception
Audiences praise the filmmaking quality and technical effort, but are significantly discouraged by prohibitive costs and accessibility barriers.
A genuinely technique-dense, tool-vendor-produced tutorial that walks through real consistency problems (facial drift, background jumping, self-healing continuity errors) and named fixes for each, though it also functions as promotion for Higgsfield's own paid platform, skill file, and Seedance model.
39:28