How to Make an AI Short Film: CROWNHELM's 13M-View Prompt

A knight stands in thick fog with two archers. Thirty metres out, a painted raider with a machete sprints straight at them. The knight raises a flat hand: hold. The archers do not understand why. Twelve seconds later a rope on the only tree in the field answers the question.

That film is CROWNHELM, Part 1. It is one Seedance 2.5 generation: 25 seconds, six cuts, four characters, sound included. It passed 13 million views, and 70% of the people who started it watched to the end.

This guide shows how it was made, with the real prompt. The prompts for this film and for the one before it, Last at the Gate, are public on their pages, so you can read every line and copy the structure.

CROWNHELM, Part 1. One Seedance 2.5 generation, 25 seconds, six cuts. Full prompt and cast.

Key Takeaways

• A 25-second film with six cuts can render as one Seedance 2.5 generation. One prompt keeps light, faces and geography consistent across the cuts.
• Faces come from character sheets. The prompt names each character by its @tag and writes in words only what the sheet does not show.
• Write the prompt in blocks: scene context, cast, location map, first frame, cuts, performance, physics, light, audio, style, locks.
• Give every distance in metres and fix the screen direction, for example "the gate is always on the left".
• Put danger in the first frame and a second event before second three.
• Read the render one frame per second and fix only the cut that failed.
• Never write "cut to black". The model renders the black inside the seconds you pay for.
• The fastest way to write and fix a scene is in chat: connect ChatGPT or Claude Code to Athana through MCP.

What you need to make an AI short film

To make an AI short film you need four things: one story beat that turns, a cast of character sheets so the faces stay the same, a scene prompt written as a cut list, and a video model that holds 25 to 30 seconds in one take. In 2026 that model is Seedance 2.5, which also generates the sound in the same pass.

The order matters. Every step below fixes a problem we hit on an earlier render.

Step 1: Pick one beat that turns

A short film is one moment where something changes. Before writing a single cut, answer three questions: whose scene is it, what does the audience know that the character does not, and what does it cost.

CROWNHELM answers all three in its first block:

This is the knight's scene. Only he knows that the rope on the tree is a snare; the archers do not, and the audience sees a rope on a tree and does not know what it is for until it fires. The camera shows his calm and one tightening hand, never his reason. If the archers loose, the man dies for nothing; if the rope fails, the machete reaches them in three strides. He bets both on a rope, and wins.

Everything the camera does later serves that paragraph. The close-up on his hand tightening on the sword hilt exists because the scene is his and his reason stays hidden.

Step 2: Build the cast before the scene

Every recurring person gets a character sheet: one image with the face, the costume and the build from several angles. The knight is @knighteyes, a sheet where the visor is closed on every panel and the eyes show through the slit. A prop gets its own sheet when its shape matters, like the kite shield in Last at the Gate.

The prompt then names each one in a cast block, one line per reference, saying who it is and what it controls:

@knighteyes is the knight, the one in command; closed visor, eyes visible through the eye opening. @archer1 is the lean archer at his LEFT shoulder. @archer2 is the broad archer at his RIGHT shoulder. @savage is the running man. Each reference controls only its own person; sheet backgrounds and layouts are not inherited.

Two rules came out of failed renders:

• Don't describe what the sheet already shows. A sentence about a face or a costume fights the sheet, and the model splits the difference. Write in words only what is on no sheet: the sword on his hip, the bows, the machete.
• The sheet must only show what should appear. An early knight sheet had one panel with the visor raised. The visor then opened in cuts where the text said it was closed. We redrew the sheet with the visor shut on every panel.

Extras get no sheet, and the prompt says so: "The ten men who appear in CUT 6 have no reference."

Step 3: Write the place in words

For the first knight film we generated a still of the valley and attached it as a location reference. It looked like a fantasy wallpaper, and that CG look spread into every cut. We dropped the still and wrote the place instead, as a map with fixed distances:

Exactly fifteen metres in front of them stands a single tall old tree, about twenty metres high: [...] Exactly fifteen metres beyond the tree, thirty metres from the knight, the grass fades into a wall of pale grey fog, and @savage comes out of it there.

Last at the Gate adds a screen rule that keeps the geography readable across cuts: "In every side view the gate is on the left of the frame and the enemy on the right."

If you do use a location photo, it has to be empty. A still with people in it makes the model copy those people, their poses and the framing, even when the text says the photo controls only the place.

Step 4: Write the prompt as a cut list

The prompt is a set of blocks. Each one answers one question for the model:

Cut lengths go only in the cut headers, as "about 5 s". Inside a cut, describe what happens and why. When we timed a fall to the tenth of a second in an earlier test, the actor fell stiffly and stared into the lens. Seedance plays a described beat more naturally than a stopwatch.

Here is the opening cut of CROWNHELM, word for word:

CUT 1, about 5 s, long lens from behind the knight's right shoulder, handheld, barely breathing: His war cry is the first sound, at 0:00.3, before anything else. @savage sprints, a full explosive sprint from the first frame: long driving strides, body leaning forward, arms pumping, the machete arm pumping with them, grass flying from his feet. [...] After two seconds two longbows rise into the frame from both edges, @archer1 on the left and @archer2 on the right, and draw to the cheek [...] and between them the knight's right gauntlet comes up flat, palm down, and stays there in the air. @archer2 whispers, off his face, "Sir?"

Three details in that cut were fixes:

• Speed is spelled out. "Runs" rendered as a jog. A sprint appeared only when the prompt said explosive start, body forward, arms pumping, and named where he is at the start and the end of the cut.
• Distances are in every cut. "He starts fifteen metres beyond the tree and by the end of the cut he is barely ten metres from it."
• The mechanism exists before it fires. In the first drafts the rope appeared only when the trap went off, and it rendered like a web shot from nowhere. The final prompt lays the rope in the grass, up the trunk and over the branch from the first frame, then fires it along that same path.

Step 5: Hook in the first three seconds

A viewer decides in about three seconds whether to stay. The first frame has to show the danger, and a second event has to land before second three. No establishing shot, no title card, nothing standing still.

CROWNHELM opens on the raider already launching out of the fog at the camera. The war cry starts at 0:00.3. The bows rise at second two. The flat hand goes up, and "Sir?" is said aloud before second five: a danger, two events and a spoken question. After that, each question stays open until the next one is already asked. Why won't he let them shoot? The trap answers at about second twelve. Then the laugh, the scream from the fog, and the line of raiders. The film ends on an open question, which is where Part 2 starts.

Step 6: Write the sound into the prompt

Seedance 2.5 generates sound with the picture, so the audio block is part of the direction. Write it per cut: footsteps, breath, the creak of a bowstring, the one spoken word. Both films use no music and a single line of dialogue, "Loose!" in one and "Sir?" in the other.

Slow motion works the same way. A whole clip written as "smooth slow motion at about half speed" came back at real walking pace. A ramp at a named event rendered exactly: real speed first, then "everything slows into slow motion, about a quarter of real speed, like a high-speed camera" as the ring falls out of the pouch, with the sound going "dull and distant, as if under water". Lock where the slow motion starts and stops, and say everything else is real speed.

Step 7: Lock each beat

The last block is a list of hard rules. Write them per beat, in plain sentences:

The knight's visor is closed in every cut; only his eyes and the skin around them show through the wide eye opening; never his whole face, never a raised visor.

Distances are fixed in every cut: the tree is fifteen metres in front of the knight, never closer.

Two locks go into every scene: "No text anywhere" and "Nobody looks into the lens".

Step 8: Render once, then read it frame by frame

Render the whole scene as one generation first. We once planned to split a 30-second, three-character scene into four parts out of caution. Rendered as one pass instead, it came out right on the first try. The first knight film went the other way: rendered scene by scene, it drifted. Fog came and went, the road turned into a meadow, and the visor opened in one scene and closed in the next. As one prompt it held.

Then read the render as a contact sheet, one frame per second. Find the cut that failed and change only that cut. A cut that rendered well is frozen: rewriting it "to make it better" usually breaks it.

The render also becomes canon. Last at the Gate asked for thirty archers on a ridge, and the model drew a whole army in the valley. The next film describes the army from the render, because that is what the audience saw.

What broke, and how we fixed it

Two ways to make one in Athana

With Mira in Film Studio. In Film Studio the steps above are the workflow. You make each character once, from a photo or a description, and tag it with @ in any scene. You describe the scene in your own words, and Mira, the AI film director, writes it in the same block structure: cast, location, cuts, performance, sound, locks. You can edit every line before you render. Seedance 2.5 is the default model, and the price shows on the Render button before any credits are spent.

With ChatGPT or Claude Code through Athana MCP. This is how our studio makes its films. Connect Athana MCP to ChatGPT, Claude, Claude Code or Cursor and sign in once with your Athana account. The agent loads Athana's director grammar, creates the characters, writes the scene and renders it on Seedance 2.5 with your plan credits, quoting the price before every render. Fixes are the fastest part. Watch the render, tell the agent which cut failed and what to change, and render again. The rest of the prompt stays as it was.

To start from these films, open Last at the Gate or CROWNHELM and press Clone to Film Studio. The project lands in your Film Studio with its prompt, so you can change the story and keep the structure.

Frequently asked questions

How long can an AI short film be?

One Seedance 2.5 generation runs from 4 to 30 seconds, with several cuts inside it. For a longer film, render scene after scene with the same cast and join them into one video. Film Studio does the joining for you.

Which AI model is best for short films?

We use Seedance 2.5 for every film in this guide. It holds up to 30 seconds and several cuts in one take, follows a long cut list closely and generates the sound in the same pass.

How do I keep the same character in every shot?

Give the character a sheet and tag it in the prompt. Write nothing in the prompt that the sheet already shows, and make sure the sheet shows only what should appear on screen.

Do I have to write prompts this long myself?

No. In Film Studio you describe the scene in plain words and Mira writes the full prompt in this structure. Through Athana MCP, ChatGPT or Claude Code writes it with you in chat, which is how our studio works. Either way you can edit every line before a render.

Can I use the knight prompts?

Yes. Both prompts are public on their film pages. Clone a film to Film Studio and swap in your own cast and story.