Source: raw/Claude_+_Seedance_2.0_is_UNSTOPPABLE_at_Video_NEW_Skill.md, raw/Claude_just_made_CINEMATIC_Ads_absolutely_easy_NEW_Skill.md (auto-caption transcripts, fetched 2026-10-27); GPT Image 2 launch date from OpenAI’s post, ai-research/introducing-chatgpt-images-2-0-openai-2026-10-27.md
Creator: Robo Nuggets (host Jay) · Platform: YouTube · Skill level: operator
Videos:
- “Claude + Seedance 2.0 is UNSTOPPABLE at Video” — youtube.com/watch?v=tnOwKpKwAZs · 17:16 · published 2026-04-14
- “Claude just made CINEMATIC Ads absolutely easy” — youtube.com/watch?v=t6w-IpKjW1s · 16:15 · published 2026-04-30
Historical context (April 2026)
Both videos predate the current wiki state. Seedance 2.5 shipped in August 2026 (see the August 2026 video-model wave), and the Claude 5.5 models and Claude Motion arrived in September and October. The Claude Desktop tabs the first video describes (Chat, Cowork, Claude Code) are also pre-merge: Chat and Cowork merged in September 2026. The workflow below still transfers: a skill holding API access, an agent writing the prompts, and a human approving each step. Re-check model names, tiers and prices before reuse.
Two April 2026 Robo Nuggets videos show Claude Code calling ByteDance’s Seedance 2.0 directly through an API aggregator, not through a Higgsfield-style subscription app. The first sets up a Seedance skill on fal.ai and uses it to make app-promo motion graphics from screenshots. The second builds that into a four-step agent: product image, marketing directions, a GPT Image 2 storyboard, then a Seedance render of the storyboard. The reusable parts are the setup pattern, the approval gates, and a few model-specific tips. Robo Nuggets sells the full agent kit to its paid community; that pitch is omitted here.
Key Takeaways
- The skill is the integration. A skill file holds two things: how to call Seedance through the aggregator, and Seedance prompting best practices. With it installed, you describe the video and Claude writes the prompt, picks resolution, duration, aspect ratio and audio, uploads the references, and polls until the clip is ready.
- Aggregators instead of a subscription.
- fal.ai: pay per use, no monthly plan, hundreds of models behind API endpoints. Jay found Seedance 2.0 most stable on fal “as of the time of this recording”.
- Kie.ai: he calls it “cheaper than most”.
- WaveSpeed.ai: a third option he uses. He says he has no affiliation with any of the three. This is the same per-generation argument as generate skill.
- API key handling. The skill writes your fal key to a
.envfile. If you would rather not paste the key into chat, ask Claude where the.envfile is and paste it there yourself. - Add a model mode by pasting its docs. Each fal model page has a “copy content” button that copies documentation written for LLMs. Paste it into Claude Code and ask it to add that capability to the skill. His example: adding Seedance image-to-video (first-frame) mode alongside reference-to-video. The same move works for any model on fal, Kie or WaveSpeed.
- Ask for three directions, then choose. Both videos request three options before any generation, for example three concepts for an app promo or three ad concepts for a product. Jay’s reason: it brings your own taste in before any credits are spent. Beginners should also watch each step rather than run it end to end.
- Seedance modes and tiers. Reference-to-video bases the clip on screenshots or product photos you supply. Text-to-video needs no assets. There is a Pro tier and a Fast tier; Jay found the price gap small and thinks Pro is worth it.
- Batch from URLs. Give Claude a list of project websites, or an App Store link, and it sources the screenshots, writes the prompts and generates one clip per project. His examples were promo intros for Project Glasswing, Karpathy’s Auto Research, Garry Tan’s GBrain, Claude’s Ultraplan and Granola.
- Limitations Jay names:
- Small text renders poorly. Large text came through; he suspects low-resolution reference images.
- Cost. A per-model cost overview Claude compiled for him put Seedance 2.0 at more than twice Kling 3.0 per second, and four to five times Veo 3.1. His estimate is $2-3 per 10-second clip. For plain B-roll he points to Kling 3.0 or Veo 3.1 instead.
- A community Seedance prompt library (name not given in the captions) can be handed to Claude with “recommend how to integrate this into our setup”, so the skill draws on proven prompts.
The four-step cinematic ad agent
Jay’s “cinematic ads” agent runs as one skill, in four steps with a human gate between each:
- Set the product. Drop in a product image plus any brief, such as “emphasize speed; end tag shows the ivory color and on.com”. The agent extracts the brand, product name and category, and checks details that were not supplied.
- Storyboard directions. It returns three distinct concepts, each with a concept name, a vibe and a product “hero moment”. For an On running shoe these were “liquid ground”, “first light sprint” and “wind tunnel”. You pick one and refine it.
- Storyboard plan, then storyboard image. The plan is timestamped scene by scene (“first two seconds: the shoe rockets towards a mint pool”) and names the camera for each shot. After approval, the agent sends a detailed prompt to GPT Image 2 (the edit endpoint, on fal) with the product photo as reference. The result is one six-panel storyboard image with correctly spelled panel text. Jay notes this step alone is useful in advertising: you can show the client a storyboard before producing anything.
- Render. Before spending money, the agent shows a confirmation card: model (Seedance 2.0 Pro), resolution, duration, aspect ratio, audio on or off, and the references it will send (the storyboard sheet and the clean product photo). The Seedance prompt itself is short, because the storyboard carries the detail: roughly “turn image 1 into a cinematic 15-second sports ad”.
He says the agent is platform-agnostic: Claude Code, OpenClaw and Hermes all work. He drives his through Telegram, connected to long-running Claude Code sessions.
The face-policy limit. Seedance tends to block reference images that contain a person’s face, which matters for ads with a human model (Jay hit it on a lip-tint ad). The video shows a way around the filter; this wiki doesn’t reproduce it, because the filter exists to prevent generating real people’s likenesses. For ads that need a person, use a model release or a licensed or consented likeness through a provider whose terms allow it, or design the shot so the product, hands or an unidentifiable figure carry it. (Edited 2026-10-27 by the coordinator.)
Try It
- Create a fal.ai key and a minimal Seedance skill. Have Claude Code write the skill from fal’s “copy content” docs for one Seedance endpoint, and store the key in
.env. - Run three directions on your own product screenshots. Pick one and generate a single Pro clip. Then ask Claude for the exact prompt it sent, so you learn what good prompts look like.
- Add a cost gate to the skill. Before any paid call, it should print the model, duration, resolution and references, and wait for your yes.
- For product ads, add the storyboard step. GPT Image 2 makes a six-panel sheet; use that sheet as Seedance’s reference image.
- Compare price before you scale. For B-roll, run the same brief on Kling or Veo, and keep Seedance for shots where its motion quality matters. On current models, start with Seedance 2.5 instead.
Implementation
Tool/Service: Claude Code (or another agent harness) with a custom skill. fal.ai endpoints (Seedance 2.0 reference-to-video and image-to-video, GPT Image 2 edit), or the equivalent endpoints on Kie.ai or WaveSpeed.ai.
Setup: Create an API key in the fal dashboard and give it to Claude Code, which stores it in .env. Paste each model page’s LLM-ready docs into the session to add endpoints.
Cost: Per generation. Jay’s figure is $2-3 per 10-second Seedance 2.0 clip, from a Claude-compiled price overview at recording time. He notes Pro costs little more than Fast.
Integration notes: Outputs come back as local MP4 paths, which other tools in the same session can use, such as a site build or a HyperFrames or Claude Motion composition. Robo Nuggets shows the results in its own console, which you do not need. A later Robo Nuggets video (October 2026) describes the reverse pairing: Claude Motion builds code-based storyboard scenes first, and Seedance turns each into a moving shot. See Claude Dashboards and Motion.
Open Questions
- GPT Image 2 launch timing. On 2026-04-30 Jay says GPT Image 2 was “released just a few days ago”. The GPT Image 2 launch article dates the launch to “~early May 2026”. Both are approximate; the exact date was not checked.
- Answered (2026-10-27): OpenAI’s post “Introducing ChatGPT Images 2.0” is dated April 21, 2026 (https://openai.com/index/introducing-chatgpt-images-2-0/), and OpenAI’s launch video was uploaded the same day. Jay’s “a few days ago” on 2026-04-30 was nine days after launch. The launch article’s “early May” was wrong and has been corrected.
- The Seedance vs Kling vs Veo price ratios came from an overview Claude produced in the video, not from a price page. Current fal pricing was not checked.
- The Seedance prompt library’s name and URL are not in the captions.
- Which video providers support consented likenesses for ads (model releases, licensed avatars) through an API Claude Code can call: not covered by these sources.
Related
- Higgsfield MCP Tutorial (Robo Nuggets) — the same creator’s Higgsfield-subscription route a few weeks later, with a fal.ai comparison
- generate Skill Instead of a Higgsfield Subscription — the pay-per-generation argument, generalized
- FREE Seedance 2.0 Claude Skill (LTX Studio) — another storyboard-grid-to-Seedance workflow
- Image Models Feed the Video Pipeline — why the GPT Image 2 storyboard comes before the video model
- GPT Image 2 — the storyboard model
- The August 2026 Video-Model Wave — Seedance 2.5, the successor
- Higgsfield + Claude Code Ad Agency Workflow — the same approval-gated ad pipeline through Higgsfield