One Idea To A Finished AI Film: The Seedance 2.5 Workflow

Most people treat AI video as a slot machine. They describe a scene, generate, change three words, generate again, and keep paying until something looks acceptable. That never adds up to a film, because the clips do not share a face, a room, or a light source. The workflow on this page treats a short film the way a production treats one: a story first, then a cast, then a world, then a storyboard you can criticise for free, and only then a render. Every stage is a place to catch a problem while it is still cheap to fix.

The worked example throughout is that film. A man hikes through fog, finds a cave, meets a yeti, gets chased, wakes up safe in bed, and then a yeti shadow creeps across his wall. Twelve beats, two characters, three locations, one storyboard image, one render.

The input to this whole pipeline is smaller than people expect. It is a premise you could say out loud in a lift. Man hikes, finds cave, yeti chases him, he wakes up, the shadow is still there. That is the entire creative brief, and everything downstream is elaboration of it.

A premise only works if it contains a change. The yeti idea works because it has a reversal built in: the danger turns out to be a dream, and then the dream turns out to maybe not be a dream. A premise with no turn gives you thirty seconds of pretty footage and no reason to watch it twice.

If you do not have an idea, ask the model to interview you instead of waiting for one. Give it your genre, your constraints, and the number of characters you are willing to build, and have it ask questions until a premise falls out. Answering five specific questions is much easier than inventing a story from nothing, and the answers become the brief.

Seedance 2.5 renders up to thirty seconds in one pass. That is the hard boundary the whole story has to live inside, and it is also the reason the beat sheet matters so much. Thirty seconds is not one shot. It is somewhere between eight and fifteen shots, and you can direct a change of shot on almost every second.

So the next artefact is a beat sheet: a numbered list where each entry is one shot with a shot size, a subject, and an action. Not a mood. A shot. The yeti film opened on a close up of boots on wet ground, cut to a medium of the hiker stopping mid stride, then an extreme wide of the cave. You can read that list and see the film in your head before a single pixel exists.

Have the model write the beat sheet from your premise, then edit it like a director. These are cheap to change now and expensive later:

• Shot variety. Four medium shots of one person in a row will feel flat no matter how good the render is.
• Screen direction. If your character runs left to right in one beat, they should not run right to left in the next unless you meant it.
• Beat count against the clock. Fifteen beats in thirty seconds is two seconds each, which is fast. Six beats is five seconds each, which is slow. Decide which your story wants.
• The reveal. Whatever the film is actually about should land in the last third, not the first.

The beat sheet does a second job that is easy to miss. It tells you exactly which assets to build. Read the yeti beats and the shopping list writes itself: one hiker, one yeti, a forest, a cave, a bedroom.

A character sheet is a reference image of one character, generated on its own, that you hand to the video model every time that character appears. It is the mechanism that stops your hero's face from changing between shots. There is a real fork here, and it changes what you do next.

• You have photos of the person. Upload three or four of them and have the sheet generated from your actual face. The tell that you are in this branch: you want yourself, a client, or a real product in the film. In the worked example the sheet was built from three photos, and it kept the face exactly while changing the wardrobe into hiking gear.
• You are inventing the character. Then you have to describe them, and vague descriptions produce a different person every run. The tell: nobody real has to be recognisable.

If you are inventing, specify these attributes explicitly rather than writing a sentence of vibes: approximate age, build and height, posture and how they carry themselves, hair, face shape, distinguishing marks like scars or tattoos, and clothing down to the footwear. The attributes you leave out are the ones that will change between shots. The yeti in the example was pinned the same way: sharp teeth, oversized hands, long claws, and those three details survived every beat.

Generate character sheets on a plain neutral grey background. This is not an aesthetic preference. A flat grey field with no directional light lets the model cleanly separate the character from the backdrop and drop them into any scene you like. The moment you generate a character sheet in a location, that location's light comes with them everywhere.

An environment plate is a still of a location with no characters in it, generated in the film's real light. Forest, cave, bedroom. Most people rush this stage, and it is the one that decides whether the film looks like a film, because the environments are where the look is set.

The test is not whether each plate looks good on its own. It is whether all of them look like they were shot by the same crew on the same production. A forest that looks like a photograph, a cave that looks like a video game, and a bedroom that looks like an advert will not cut together. When you describe a location, answer four questions rather than writing an adjective pile:

1. 1What is the space, physically? Architecture, materials, surfaces, what is on the walls and the floor.
2. 2What is lighting it? Windows, lamps, screens, fire, an open cave mouth. Name the actual source.
3. 3What time and weather is it? This is what makes three separate locations feel like one continuous afternoon.
4. 4What is the overall look? Photoreal film, animation, illustration. Say it out loud rather than leaving it to chance.

That fourth question is the one people cannot answer, and it is why the next stage exists.

This is the single most useful move in the whole workflow, and it generalises far past AI video. Most people cannot describe the look they want, because the vocabulary is a professional one. They know a frame looks beautiful. They cannot say it is a fine grain stock with warm practicals and a cool ambient fill. So they type cinematic and get the model's average idea of cinematic, which is the same idea everyone else is getting.

The fix is to stop describing and start extracting. Find a still from a film whose photography you admire, screenshot it, and hand the image to the model with an instruction to analyse it and write back one paragraph describing how it was captured. You are not asking for an opinion about the image. You are asking for a specification of it. Have it define the capture medium and format, the grain, the quality and direction of the light, what the camera is doing, the lens character, and the colour treatment.

In the worked example the reference was a grid of Lord of the Rings frames pulled off a normal image search. What came back was a paragraph of proper cinematography terms about the stock, the negative, and the colour. You do not need to understand the words for them to work. That paragraph gets pasted back in as the look for every environment plate, and the second pass went from looking like AI images to looking like frames from a film.

A look paragraph is reusable. Build two or three from references you like and you have a small library of house styles that are yours rather than the model's default.

You now have a beat sheet, characters, environments, and a look. The temptation is to start generating video. Do not. Generate the storyboard first, and generate it as a single image containing every beat as a panel.

This is a pricing trick as much as a creative one. Nine or twelve panels rendered as one image costs a couple of credits. The same twelve beats rendered as video cost many times that, and if the story is wrong you pay it again. The storyboard is the cheapest possible version of the finished film, and its only job is to let you find out that something is wrong.

Read the board as a whole rather than admiring individual panels. You are checking whether the story reads without words, whether the reveal lands, and whether a stranger could follow it. The yeti board showed the hike, the spot, the climb, the entry, an enormous eye in the dark, the encounter, the chase, and the wake up, and it was legible end to end at a glance.

Storyboards exist to catch the errors that only appear when shots sit next to each other. The yeti film had exactly one. Beat 11 had the hero running towards camera, terrified, fleeing the cave. Beat 12 had him facing the yeti's claw, which meant he had turned around between two consecutive shots. Each panel was fine. Together they broke the chase, because the audience would lose track of which way anyone was moving. The note back to the model was specific: keep the hero facing camera as he runs towards us, with the claw reaching towards his back, so beat 12 matches beat 11. The revised board fixed it in one pass.

That note is worth studying, because vague feedback produces vague revisions. It names the beat, the problem, the fix, and the reason. Notes like make it scarier will get you a different image, not a better one.

• Direction breaks. A character or a threat that flips sides between adjacent beats.
• Wardrobe and prop drift. A jacket, a bag, or a weapon that appears or vanishes.
• Scale drift. The thing that is meant to be enormous quietly shrinking to human size.
• Time of day drift. Fog in one beat, clear sun in the next, no reason given.
• Repeated framing. Two beats that are effectively the same shot, which wastes seconds you do not have.

The final prompt is not a description of a scene. It is a production brief. Accuracy jumps when you write it this way because you are giving the model the same information a human director would need to shoot the thing. Nine blocks do the work.

• Premise. One line on what the film is and what genre it belongs to.
• Arc. What changes from the first second to the last, so the film has a shape rather than a sequence.
• Assets. Which character sheets and environment plates exist and which beats each one belongs in.
• Invariants. The things that must never change. This block earns its place: in an earlier film the note that the dragon must always be larger than the house is what stopped the dragon quietly scaling down every time it appeared.
• Beats. The timed shot list, second by second.
• Look. The cinematography paragraph from stage 5, pasted in whole.
• Direction and timing. Camera behaviour and how long each shot holds.
• Sound. What the audience hears, including silence. A film with no dialogue is carried entirely by faces, actions, and sound design.
• Exclusions. What must not appear. Cheaper to forbid than to regenerate.

The invariants block is the one most people skip and the one that saves the most money. Everything else describes what you want. Invariants describe what the model is not allowed to drift on, and drift is what ruins multi shot sequences.

You can write all nine blocks by hand and the film will come out well. The installable skill that assembles them for you, along with the exact prompt wording for the look extraction and the asset stages, lives inside the club at https://www.skool.com/claudecodeclub/about.

One more gate before the real render, and it is the cheapest insurance in the pipeline. Render at 480p first. It will look soft and small and nobody else should watch it. It exists so you can check flow: does the pacing work, do the cuts land, does the reveal read.

720p costs roughly twice as much as the test. So the decision is straightforward, and it has clear tells on both sides. Skip the test when the storyboard is very simple, the beat count is low, and you would rather have the finished file now than save a little. That is the branch taken on camera in the walkthrough, for time. Run the test when the film has more than about eight beats, when timing carries the joke or the scare, when a client is paying, or when you are going to iterate on the brief anyway. If you plan to change anything at all after seeing it, test first, because the second render is the one you were always going to pay for.

Five failures account for nearly everything that goes wrong, and all five are upstream of the render.

• Generating video before the storyboard. Every story problem costs many times more to find in video than on a board. This is the expensive mistake and the most common one.
• Character sheets with a look baked in. Neutral grey and flat light, always. A sheet with mood in it poisons every scene that uses it.
• Environments built one at a time with no shared look. Write the look paragraph once, apply it to all of them, and regenerate the odd one out.
• A premise with no change in it. No production quality rescues thirty seconds in which nothing happens.
• Vague revision notes. Name the beat, the problem, the fix, the reason. Anything less is rerolling.

The shortest honest path from finishing this page to a finished short. Expect one to two hours the first time, and much less after that.

1. 1Write your premise as one sentence containing a change. If it has no turn in it, rewrite it before going further.
2. 2Have the model expand it into a numbered beat sheet, one shot per entry, sized to thirty seconds. Edit it for shot variety and screen direction.
3. 3Read the beat sheet and write down every character and every location it needs. That is your asset list, and it is complete.
4. 4Find a still from a film whose photography you want, and have the model analyse it into one paragraph specifying capture, grain, light, camera, lens, and colour. Save that paragraph.
5. 5Generate a character sheet for every recurring character on plain neutral grey. Use photos if the character is a real person, or a full attribute list if they are invented.
6. 6Generate an environment plate for every location, all of them using the same look paragraph. Reject any plate that does not belong to the same world as the others.
7. 7Generate the entire storyboard as one image with a panel per beat. Read it end to end for story, not for beauty.
8. 8Write revision notes that name the beat, the problem, the fix, and the reason. Regenerate the board until the story is right.
9. 9Assemble the nine block brief: premise, arc, assets, invariants, beats, look, direction and timing, sound, exclusions.
10. 10Render at 480p, watch it once for flow, apply any last notes, then render the final at 720p.

Do this twice and the second film is much faster than the first, because your characters, your environments, and your look paragraph are all still sitting there. The pipeline is the asset, not the film.