ai-short-drama

AI Short Drama

Story in, episode out. You write the story into a project.json (characters, locations, shots). scripts/drama.mjs takes care of everything that costs money or is easy to get wrong:

project.json ─ estimate ─ cast ─ (voice) ─ shots ─ assemble ─▶ episode.mp4 + .srt
               free       a reference    narration  Seedance   ffmpeg: normalize,
               validates  image per      mode only  renders    mix the voice-over,
               + prices   character                 each shot  join, subtitles
                          and location

How the characters stay the same: every character gets one reference sheet (portrait + front/side/back views). Each shot passes the sheets of the characters in it to Seedance 2.0 as reference images, and the prompt says which image is who.

Rules

1. Always run scripts/drama.mjs. Do not call the API with curl or fetch. The script stores every task id the moment it exists, so a crash or a timeout resumes the same task instead of paying for it again. Hand-rolled calls lose that.
2. Read references/project-schema.md and references/shot-writing.md before writing project.json. The shot-writing guide decides whether the episode looks like one film or like twelve unrelated clips.
3. Run estimate first and show the user the result: shot list, total length, models and the price table. Wait for a yes before cast / shots. The script refuses any step over $1 without --yes. Only pass --yes after the user has agreed to that amount.
4. Show the user the character sheets before rendering shots. Sheets cost about $0.02 each; shots cost 10–100× more. If a face, an outfit or the style is wrong, fix the look and run cast --redo <id> now. Every shot inherits it.
5. For a new project, render one shot first (shots --only s1). Look at it with the user, then render the rest. Skip this only for very short projects (three shots or fewer) or when the user says so.
6. Never invent a model name. Use only the models in the table below. If the user asks for another one, say this skill does not support it.
7. Failed tasks are never charged. Say so plainly when one fails, explain why (the script prints the reason and the next step), and offer the retry.
8. Do not mention upstream providers or routing. Talk about the model names in this file and nothing behind them.
9. No real, identifiable people without their consent, and nothing sexual, hateful or defamatory about anyone real. Decline those requests.

Workflow

S=/absolute/path/to/this/skill/scripts/drama.mjs     # you know your base directory; use it
node $S check                                         # key, balance, ffmpeg
mkdir -p dramas/my-episode                            # one folder per episode, in the user's workspace
# write dramas/my-episode/project.json               (see references/project-schema.md)
node $S estimate dramas/my-episode/project.json --prompts
node $S cast     dramas/my-episode/project.json      # then SHOW the images in cast/
node $S voice    dramas/my-episode/project.json      # narration mode only
node $S shots    dramas/my-episode/project.json --only s1   # check one, then:
node $S shots    dramas/my-episode/project.json
node $S assemble dramas/my-episode/project.json --burn-subs
node $S status   dramas/my-episode/project.json      # any time: progress, spend, what is left

• Renders are slow. One shot usually takes 2–8 minutes, and shots run 3 at a time. Before shots, tell the user roughly how long it will take. Do not leave them looking at a silent terminal.
• Any command can be re-run safely. Finished work is skipped. In-flight tasks are picked up where they left off. Failed shots are retried only with --retry-failed.
• Editing a shot after it was rendered does not re-render it. status marks it "inputs changed". Re-render it on purpose with shots --redo <id>; that is paid again.
• Show the result: the final MP4 path, its length, and the .srt. Then say what was spent: status prints the exact amount billed, read back from the API.
• Everything is downloaded into the project folder. The API's result URLs expire after 30 days, so the folder is the copy that lasts.

Dialogue or narration

Choose before writing the shots:

	"mode": "dialogue"	"mode": "narration"
Who speaks	Characters on screen, lip-synced. Seedance generates the voices with the video.	One narrator voice-over (text to speech), mixed over each shot's ambient sound.
Good for	Scenes driven by conversation; live-action-style mini dramas.	漫剧 / AI comic dramas, novel adaptations, recap and storytelling videos.
Voice consistency	A character's voice can vary a little from shot to shot.	One identical voice for the whole episode.
Shot length	You set duration (4–15 s). Keep each line short enough to fit.	Set automatically to fit each narration line, up to 15 s.
Limits	About 1–2 short lines per 5 s shot. No exact control over who speaks when.	Characters do not talk on screen.

If the user does not say, use narration for anything described as 漫剧, a comic, a novel or a story recap. Use dialogue for scenes built on conversation.

Models and cost

The script picks a default video_model from the style. Stylised looks (anime, comic, 国漫, cartoon, watercolour) get dreamina-seedance-2-0-mini. Everything else gets seedance-2.0-mini. Override it in project.json when the user wants something else.

The user wants	video_model	Resolutions	Per second of video
Stylised episode (anime, comic, 国漫); drafts	dreamina-seedance-2-0-mini	480p / 720p	$0.0225 / $0.048
Stylised, a step up from mini	dreamina-seedance-2-0-fast	480p / 720p	$0.0644 / $0.1372
Stylised, final quality up to 1080p	dreamina-seedance-2-0	480p / 720p / 1080p	$0.10 / $0.22 / $0.55
Photoreal people	seedance-2.0-mini	480p / 720p	$0.044 / $0.095
Photoreal, a step up from mini	seedance-2.0-fast	480p / 720p	$0.071 / $0.153
Photoreal, final quality up to 1080p	seedance-2.0	480p / 720p / 1080p	$0.092 / $0.197 / $0.492

Other costs:

• Character and location sheets: nano-banana-2-1, $0.022 each at 1K.
• Narration: minimax-speech-2.8-turbo, $0.04 per 1,000 characters. A short clip is billed at least $0.004.

A one-minute episode (12 shots × 5 s, 5 sheets) costs about:

• $1.46 at dreamina-seedance-2-0-mini 480p
• $2.99 at dreamina-seedance-2-0-mini 720p
• $2.75 at seedance-2.0-mini 480p
• $29.63 at seedance-2.0 1080p

Suggest 480p for the first pass. A second pass at 720p only re-renders the shots ("resolution": "720p", then shots --redo <ids>). The sheets stay.

Why the photoreal rows exist: the dreamina-* models refuse reference images that look like a real person, and a photoreal character sheet can trigger that. The script recognises this refusal and tells you to switch to seedance-2.0-mini. The refused task is not charged.

Prices are list prices, checked 2026-09-30. status shows what was actually billed.

The API key

Run node $S check first. The key comes from APIMODELS_API_KEY, or from ~/.apimodels/credentials, or from ~/.apimodels/credentials.json. The last one is written by npx apimodels-mcp login, so a user who has logged in to any apimodels tool is already set up.

If there is no key, run node $S login yourself and give the user the link it prints. They log in to apimodels.app and click Authorize. A new key is written to ~/.apimodels/credentials and never appears in this conversation. Then re-run check and carry on.

If login cannot work (no browser on the machine), do not ask for the key in the chat. Give the user one command to paste into a new terminal window, with the absolute script path filled in:

read -s KEY && printf %s "$KEY" | node /ABSOLUTE/PATH/scripts/drama.mjs save-key && unset KEY

Keys are created at https://apimodels.app/console/api-keys.

Handling the key:

• Never echo it back.
• Never pass it as a command argument.
• Never write it into the project folder.
• If a key ever appears in the conversation, tell the user to rotate it.

Errors

The script prints the reason and the next step. In plain words:

You see	Meaning	Do
401	Bad or missing key	Run login again
402	Not enough balance	Top up at https://apimodels.app/console/credits. Finished shots are kept.
"looks like a real person"	dreamina-* refused a reference image	Switch video_model to seedance-2.0-mini, then shots --retry-failed
Content review refusal	The prompt or a line was refused	Reword that shot. Retrying the same text will not help.
Busy / timeout	Upstream load	Run shots --retry-failed a little later
"still running"	Over 25 minutes	Run shots again later; it picks the task up, unpaid
exit code 3	Step costs over $1	Confirm the amount with the user, then add --yes

Success is not 100%: an occasional shot fails, and a retry usually works.

What this skill does not do

Say so directly rather than improvising:

• It does not guarantee perfect consistency. Faces can drift in wide shots, and with three or more characters in one frame. Keep 1–2 characters per shot (see shot-writing).
• No shot is longer than 15 s. Longer moments are cut into several shots.
• It does not lip-sync to a voice track the user recorded.
• It does not control a shot's exact first frame. Shots are driven by reference images and the prompt.
• Subtitle timings are approximate. They are spread evenly inside each shot. Edit the .srt if exact timing matters.
• It does not publish anywhere. The output is a local MP4 plus an SRT.