← Back to Pulse PULSE. brief
Seedance 2.5 is UNSTOPPABLE for Cinematic AI Filmmaking!

Seedance 2.5 is UNSTOPPABLE for Cinematic AI Filmmaking!

Duncan Rogoff | Learn Claude Code11 min2026-08-09 ▶ Watch on YouTube
ℹ️
Partly verifiedA few specific details here couldn't be independently confirmed against the video. The overall summary is sound, but double-check exact numbers or names before you rely on them.
What this video is
⚡ a 12-minute video, readable in 60 seconds

This video is a demo/tutorial, not a system the creator shows himself using day to day. Duncan Rogoff, a former art director for Apple, PlayStation, and Nissan, walks through building a 30-second sci-fi love-story short film by using Claude (web, Opus 5 High) to brainstorm the concept and write character-sheet, food-tray, and environment-plate prompts, Higgsfield.ai's GPT Image 2 to render the character, prop, and location images, and Seedance 2.5 to animate the assembled images into clips up to 30 seconds long. Partway through he demonstrates a more automated version of the same pipeline built as Claude Code skills, including /seedance-asset and /seedance-scene, which interview the user for an asset type, subject description, project slug, and aspect ratio, then call Higgsfield automatically via MCP; he runs this once on a test character and shows the resulting image land in Higgsfield's Assets section. The finished prompts and skills are also sold as a paid Claude skills pack linked in the video description, which is commercial framing rather than a neutral how-to.

System does [01:19]: Creator builds a 30-second cinematic AI short film, a sci-fi love story between a man and a humanoid robot lunch lady on a spaceship.
Key takeaways
+ 75 more takeaways
  • Stack [01:15]: Claude on the web (Opus 5 High) was used to brainstorm five ideas for the 30-second cinematic short.
  • Stack [01:32]: Rogoff picked concept #4, 'The Long Shift,' because aging the male character over the story gave more creative options.
  • Stack [01:55]: Higgsfield.ai's Image tab with the GPT Image 2 model was used to generate character images.
  • Stack [02:06]: Video models reportedly work best with neutral gray backgrounds for character images.
  • Stack [06:56]: A separate storyline document titled 'Spaceship lunch lady romance' held five sci-fi romance concepts each built to land a turn in under 30 seconds; the presenter picked concept #3 there as the cleanest three-act shape.
  • Stack [07:59]: Seedance 2.5 clips can run up to 30 seconds.
  • Stack [08:07]: Seedance 2.5 currently outputs only 720p; 1080p and 4K are said to be coming soon.
  • Stack [08:12]: Seedance 2.5 can be loaded with up to 50 reference images.
  • Note: extracted points describe prompt contents rather than quoting on-screen prompt text verbatim, so the lines below summarize prompt specs rather than transcribing them.
  • Prompt spec [02:13]: The four-panel character sheet prompt's count sentence at the top is load-bearing; dropping it makes the model quietly return three panels instead of four.
  • Prompt spec [02:14]: A 'Fill-In Rule' instructs leaving blank anything unknown and letting Claude work it out from the rest of the prompt without asking first.
  • Prompt spec [03:29]: The generated prompt specifies a four-panel character sheet (full-body front, three-quarter, back, and tight head-and-shoulders close-up), all showing the same person.
  • Prompt spec [03:29]: Hero character described as a 35-year-old man, athletic-lean, roughly 180cm tall, short dark brown hair, blue-grey eyes, a beard with a grey chin patch, rectangular black glasses, and a silver stud earring.
  • Prompt spec [03:29]: Hero's clothing specified as a two-piece silver reflective suit in brushed metallic finish with matching silver reflective ankle boots, worn open over a plain crewneck base layer.
  • Prompt spec [04:56]: The aged (60s) hero prompt describes a 64-year-old man, roughly 179cm tall, thinned silver-grey hair, blue-grey eyes, a long oval face, and a short nearly-white beard and mustache, with rectangular black glasses.
  • Prompt spec [05:27]: The presenter revised the lunch lady prompt so she reads as 'a true robot made of metal with distinct panels, while still looking beautiful.'
  • Prompt spec: The robot server's character sheet locks in brushed steel panelling, pink glass eyes, a white mesh hairnet over a metal scale cap, a mint work dress, a white bib apron, and white nitrile gloves (no source timestamp given).
  • Prompt spec: The tray character sheet specifies six wells with an amber protein loaf, jade cubes, a beige grain brick, a violet gel, and coral pearls, with a spork in the edge slot (no source timestamp given).
  • Prompt spec [06:00]: A rule for the product sheet template states Panel 1 must be the flat orthographic master view, never the three-quarter, and any hidden surface needs its own full description.
  • Prompt spec [06:29]: The environment plate is the last asset piece, described as the only asset that carries the film's look.
  • Prompt spec [06:47]: The environment plate must describe the actual location, architecture, materials and surfaces, what's on the walls, what the camera can see, and the camera and lens type.
  • Prompt spec [07:22]: The exterior ship Look spec calls for 65mm anamorphic on large format digital, 2.39:1 aspect ratio, deep focus at f/5.6.
  • Prompt spec [07:22]: The exterior Look also specifies cold, desaturated color with warmth held only in ivory viewports and two red position lights, fine natural grain, and hard directional lighting from a single distant source.
  • Prompt spec [07:26]: The food tray product sheet requires a flat overhead top view, a three-quarter view, an underside view, and a tight macro close-up, all of the same single loaded cafeteria meal tray.
  • Prompt spec [07:39]: Claude wrote both Look blocks cold and desaturated rather than warm tungsten, reasoning that the story's warmth is one person, not the room.
  • Prompt spec [07:39]: Claude flagged that it placed a tray on the table in the 'empty' interior scene because the story lives on that seat, offering to cut it if a truly bare room was wanted.
  • Prompt spec [08:39]: The Scene Prompt template states PREMISE, ASSETS, and LOOK are mandatory blocks; every other block must earn its place and be cut if not needed.
  • Prompt spec [08:39]: The scene-prompt fill-in rule again says leave unknown fields blank and let the model work them out itself without asking first, then report in two lines what was decided.
  • Prompt spec [08:41]: The ASSETS section requires specifying character face, build, and wardrobe exactly per reference, explicitly not using the reference sheet's grey background or flat lighting.
  • Prompt spec [08:45]: The LOCKS section fixes one location, one lighting setup, and who stays frame-left versus frame-right; the BEATS section is timestamped in seconds and each beat ends with 'NO CUT' or 'CUT.'
  • Prompt spec [08:45]: A spoken line is estimated to run 2-3 seconds, so dialogue length is counted first when building beats.
  • Prompt spec [09:18]: Prompting is described as needing to cover the location, who's in it, what they're wearing, the props, the dialogue, the overall look, and how often to cut.
  • Prompt spec [09:39]: Stated final premise: a man collects the same tray of food from the same robot server every day of a decades-long voyage, aging 30 years while she stays exactly the same, until on the final day he can barely lift the tray and she comes around the counter to dance with him instead, running 32 seconds.
  • Prompt spec [06:20]: An earlier framing of the same idea said the man would get the same tray of food from the robot for about 40 years.
  • Prompt spec [10:03]: The finished scene is broken into 15 shots with timecoded beats (e.g. 0-2s, 2-4s, 4-6s), shot on a 40mm large-format camera look.
  • Prompt spec [10:03]: Beat 11 (21-24s) shows him trying to lift the tray with shaking hands before lowering it back; Beat 12 (24-27s) has the robot leave the tray and walk around the counter to hold out her hand; Beat 13 (27-30s) has them join hands and begin turning; Beat 14 (30-32s) holds them still turning with the tray left unclaimed.
  • Prompt spec [10:03]: The dance sequence specifies no handheld camera, no shake, no zooms, no whip pans, no drone moves, and no dolly push, with the camera locked or moving only very slightly.
  • Prompt spec [10:13]: The negative prompt excludes brand logos, extra people, on-screen text, other diners, background aging effects on the robot, glowing panels or seams other than her eyes, lens flare, music sting, and a kiss.
  • Wiring [02:43]: The presenter copies Claude-generated prompts and pastes them into a new plain Claude chat (not Cowork or Claude Code) to fill in character details.
  • Wiring [03:09]: Personal photos can be uploaded into Claude so it has a sense of the user's own appearance for character-sheet generation.
  • Wiring [03:49]: A personal reference photo is uploaded before generating the output in Higgsfield.
  • Wiring [04:23]: For the robot character, the presenter removed the personal reference image in Higgsfield so the generated character would not resemble the user.
  • Wiring [05:00]: The previously generated character sheet was uploaded as a reference image so the aged (60s) hero would wear the same outfit.
  • Wiring [05:04]: Image generation was done at higgsfield.ai/ai/image using the gpt_image_2 model.
  • Wiring [05:18]: The presenter copied the second (80s) prompt and generated it using the same reference image.
  • Wiring [05:48]: The creator identified two more shots still needed: a close-up of the food tray and an exterior spaceship shot.
  • Wiring [05:56]: An interior cafeteria shot was also flagged as needed since that's where the story takes place.
  • Wiring [06:07]: The creator copied the product-sheet-style prompt template to generate a tray of futuristic sci-fi cafeteria food in close-up.
  • Wiring [07:12]: The first requested environment prompt was an exterior shot of the entire spaceship.
  • Wiring [07:17]: The second requested environment prompt was the interior of a futuristic cafeteria.
  • Wiring [07:31]: Claude pasted in the prompt for the exterior of the ship next.
  • Wiring [07:36]: The interior cafeteria prompt was described as arguably the most important prompt.
  • Wiring [07:41]: The presenter increased the batch size to generate a couple of different images so he could pick his favorite.
  • Wiring [07:53]: If you want to change an image, you can just tell Claude what you want to fix.
  • Wiring [07:57]: After downloading both the exterior and interior images, the narrator said they could now create the entire story.
  • Wiring [08:38]: The creator built a prompt template designed to reference all provided images and break the scene into separate beats.
  • Wiring [08:47]: The creator copied the filled template and pasted it into Claude to generate the finished scene prompt.
  • Wiring [08:53]: The creator planned to replace all character references in the prompt with '@image 1', '@image 2', '@image 3', etc.
  • Wiring [09:27]: The creator's reasoning for giving the model the premise is that the models are smart enough to use it.
  • Wiring [09:35]: The overall story was given to the model so it would know how to craft the beats end to end.
  • Wiring [10:20]: The full scene prompt was pasted into Seedance 2.5, which automatically tagged all previously uploaded image assets as references before generating.
  • Wiring [10:27]: Duncan turned this process into Claude Code skills that design characters, props, and environments and create Seedance clips almost fully on autopilot.
  • Wiring [10:38]: Typing the slash command for a Seedance asset and hitting enter triggers Claude to interview the user about what they're trying to make.
  • Stack: The shown Claude Code skill list includes /seedance-asset, /seedance-scene, /breakdown-radar, /yt-cinematic-intro, /ig-channel-remix, /motion-graphics, /viral-remix, /higgsfield-generate, and /hyperframes among others (no source timestamp given).
  • Wiring: The /seedance-asset skill builds reusable production assets on Higgsfield (character sheets, product sheets, environment plates) and saves each as a Higgsfield Element taggable as @name, with one hard gate: sheet approval before the Element is saved (no source timestamp given).
  • Wiring [10:50]: The interview asks the user to state the asset type (character, product, or environment), describe the subject in a sentence, and give a project slug plus aspect ratio (default 16:9).
  • Demo result [10:56]: Duncan tested the skill with the description 'a beautiful blonde woman in a red dress with bright green eyes.'
  • Wiring [11:01]: Claude calls Higgsfield automatically via MCP to generate the character image.
  • Wiring [11:07]: The resulting character image appeared in Higgsfield's Assets section once generation finished.
  • Output shown [11:38]: A man in a shiny silver blazer sits alone in the spacecraft cafeteria examining a small round food item with a compartmented tray in front of him.
  • Output shown [11:43]: A humanoid robot in a maid/server uniform with apron and cap walks past a stainless steel cafeteria serving line inside the ship.
  • Output shown [11:46]: An older man in a light gray suit slow-dances with the humanoid maid robot in the ship corridor/cafeteria set.
  • Output shown [11:50]: The humanoid robot stands at attention behind the serving counter facing the camera with the food tray placed in front of her.
  • Cost [03:56]: Generating more images in a batch costs more Higgsfield credits.
  • Limitation [05:07]: An early generated lunch-lady image just looked like a human rather than a robot, which the presenter attributed to not reading the prompt closely enough, requiring a return to Claude to fix it.
How this brief was shaped: Personal AI Build / System Demo · confidence Medium

Creator walks through a multi-tool AI stack (Claude for story ideation, an image tool for character sheets, a video model for the cinematic output) and OCR shows a dense verbatim character-sheet prompt being built on screen, which matches a system/prompt rebuild more than a generic tutorial.

The lens sets this brief's structure, never its facts — every claim is held to the same citation and fact-check standard.

Jump to a moment
0:00- Seedance 2.5 first look0:48- What you'll build today1:06- How to build your storyline1:37- How to set up Higgsfield1:55- How to create character sheets2:38- How to prompt your hero3:56- How to create robot characters4:37- How to age your characters5:52- How to create prop sheets6:25- How to build environment plates7:59- Seedance 2.5 features explained8:17- How to prompt multi-shot sequences9:05- Cinematic prompting tips explained10:25- Claude Skills on autopilot11:11- Watch the finished film
Their links, sorted & clickable
🏛️ Communities & courses1👉skool.com
🔗 Other links1👉higgsfield.ai
← Back to Pulse Dashboard
Was this brief useful?