How to Build an AI Short Film, Animation, and Video Workflow — From Design and Characters to Finished Film

Think in Shots First, Film Second

The most common beginner failure is cramming thirty seconds of story into one ten-second prompt: the character has to stand up, walk to the window, pull back the curtains, turn around, and deliver a line — and the model executes one or two of those vaguely while the rest dissolves into glitchy nonsense. Video models hold far less than most people expect:

• 5 seconds ≈ 1 complete action; 10 seconds ≈ 2 consecutive actions. That's the ceiling Seedance's official guide states outright, and our testing puts other models in the same range. Overloaded actions don't get "compressed" — they get randomly dropped.
• One camera move per shot. Stacked instructions like "dolly in while orbiting, then tilt up" are the number-one source of unstable footage. For complex movement, split into two shots and cut.
• Above 10 seconds, use a time-segmented structure: "0-3s: …; 4-8s: …" — each segment carrying its own action and camera move. It's far more controllable than one long paragraph of prose.
• Shot density has official reference numbers (the Kling 3.0 guide): 1-2 storyboard beats in 5 seconds, 2-4 in 10, 3-6 in 15. Do the math against that table before you write the shot script.

The Three Places Prompts Fail Most

1. Negation barely works. Most video models have no negative-prompt parameter, handle "no X" unreliably, and can even trigger the opposite — you write "no camera shake," it reads "shake." The fix is rewriting everything as a positive alternative: "no shaking" → "locked-off camera"; "no clutter in the background" → "a clean, empty, solid-color background." And anything a parameter can handle should stay out of the prompt entirely (a locked-off camera should go through the camera_fixed parameter). Kling 3O is one of the few models with native in-text negation support — but even there, we back up high-risk negations (like "no subtitles," which has a history of firing in reverse) with positive phrasing.

2. Multiple subjects need attributes bound one by one. "A man in red and a woman arguing in the rain" — what is the woman wearing? The model will decide for you, differently every time. Write "a man in a red trench coat, and a woman in a beige dress": every attribute of every subject explicitly adjacent, nothing left for the model to guess.

3. Longer is not better. Seedance's official guide recommends keeping prompts under 500 Chinese characters — or roughly 1,000 words in English: past that, information scatters, the model grabs the highlights and drops the details, and elements go missing. Our experience says err on the lean side — an 800-character prompt stuffed with adjectives usually loses to a 300-character version where every sentence is a concrete visual instruction. Also: "left" and "right" always mean screen left and right (the viewer's perspective). Write "left side of the frame," "foreground/background" — never "on his left."

Dialogue and Sound: A/V Sync Is Written, Not Hoped For

Today's leading video models (Seedance 2.0, Kling 3O) can generate shots with spoken dialogue directly, but there are craft rules:

• Put dialogue in quotation marks: The man calls after the woman: "Remember — never point your finger at the moon." What's inside the quotes is what gets spoken, not part of the prompt — and in multilingual prompts, keep dialogue in its original language.
• Dialogue length must match shot duration. A 5-second shot can't hold a 40-word line; the result is drifting sync or chipmunk-speed speech. Read the line aloud and time it before setting the shot length.
• Action first, then the line (visual anchoring): "The agent in black slams the table. [Agent in black, shouting in fury]: 'Where is the truth?'" — with the action first, the model knows what the speaker is doing while talking.
• One label per character, all the way through — no pronouns. In a multi-shot script, call him "the agent in black" in shot one and "he" in shot two, and the character may change faces. Repeat the same label verbatim to the end.
• Describe sound effects and BGM in plain sentences ("the rain grows heavier; a bell tolls in the distance"). Note that Seedance outputs mono — do your mix in post.

Models Speak Dialects: One Storyboard, Three Phrasings

Each model wants different phrasing — which is why we write a separate prompt skill card for every model we integrate:

• Seedance 2.0 wants long narrative sentences, written the way you'd read a storyboard to a cinematographer: what's in the frame, then how it moves, then how the camera covers it. Chinese-native — Chinese quality matches English.
• Kling 3O is a multimodal orchestrator; prompts read more like director's instructions: which asset does what must be cited explicitly ("make the character from [image 1] wave at the camera"). Handing it images without declaring their purpose is the number-one failure mode in its official docs.
• Veo and the Midjourney family are noticeably steadier in English. And regardless of vendor, camera terminology is more reliable in English (dolly in, tracking shot, POV) — it's the common denominator of every model's training data.
• Image reference modes are mutually exclusive: first-frame, first/last-frame, and multimodal reference are three distinct modes that can't be mixed. To join two shots seamlessly, use the previous shot's last frame as the next shot's first frame.
• When face consistency is weak, give multi-angle references: front + profile + 45 degrees holds up across shots far better than a single frontal photo. Also note that Seedance's multimodal reference mode officially does not support photorealistic human face material.

Pipeline: Making Shot Twenty as Steady as Shot One

Craft at the single-shot level doesn't solve the whole-film problem. Around shot ten, the real enemy is scatter: the character reference is 40 turns deep in a chat log, the camera phrasing that worked is in another window, and nobody remembers why the last version got rejected. We organized MajoFlow's orchestration canvas as "script → assets → storyboard → shots → video → review," which is really just the craft rules above frozen into a pipeline: character sheets live in a character pack referenced by every shot, winning prompts and parameters accumulate in a shot library, and every generation records which references it used and why that version won. None of this strictly requires a tool — a rigorously maintained folder and spreadsheet will do — but someone has to do it, or shot twenty will inevitably come out worse than shot one.

FAQ