A Claude skill can turn a rough idea into a tight Seedance 2.5 prompt and send it to a video model, so you stay in one chat from brief to finished clip. This article shows a working SKILL.md, the five part prompt formula, audio and dialogue rules, and fixes for clips that miss.
Most people meet Seedance 2.5 the slow way. They open a video tool, type one vague sentence, and get a clip that looks nothing like the idea in their head. The faster route runs through Claude. A Claude Seedance 2.5 skill is a small instruction file that teaches Claude how to turn a rough idea into a tight, shot-ready prompt, then sends that prompt to a video model and waits for the result. You stay in one conversation from first thought to finished MP4. This article builds that skill piece by piece, shows the prompt formula it follows, and lists the audio, image-to-video, and troubleshooting rules that decide whether a clip lands on the first try or the fifth.
There is no button labeled "Seedance skill" inside Claude. The idea is a pattern with two halves that work together: a skill that writes the prompt, and a connector that renders it.
A skill is a folder with a SKILL.md file in it. The top of the file holds a name and a short description. The rest holds instructions in plain language. Claude reads the description, decides whether your request matches it, and only then loads the full instructions. That matters for video work because the instructions run long: formulas, vocabulary lists, checklists. None of it clutters the chat until you ask for a clip.
The skill never renders a single pixel. Rendering happens on a GPU behind a connector, and the PicassoIA connector for Claude exposes everything the skill needs: tools for image generation, image editing, two video models, and job management. Under the hood it speaks MCP, the protocol Claude uses to call outside tools. The two video models are PicassoIA Video, which generates from text or an image, and Seedance, which adds synchronized audio.
Video jobs are asynchronous. The tool returns a prediction ID and an estimated time right away. Claude then checks the status tool after the suggested wait, and keeps checking until the job reads succeeded or failed. A failure is final, so the skill tells Claude to submit a fresh generation instead of retrying the same ID.
| Piece | What it does | Where it lives |
|---|---|---|
| Skill file | Rules for writing the prompt | Claude |
| Connector | Sends the job and reports status | PicassoIA |
| Video model | Renders picture and sound | PicassoIA GPUs |
💡 The connector lists Seedance 2.5 Lite as its Seedance model: 480p or 720p, 5 or 10 seconds, audio on by default. The full Seedance 2.5 page on the site goes further, with clips up to 30 seconds and up to 30 reference images. Check which one a tool call reaches before you promise a client a clip length.
Seedance responds best to cinematic, chronological descriptions: who is in frame, what moves, how the camera moves, and what the light does, in the order it happens. That is a shot list, and a language model writes shot lists well once it has a format to follow. Without the skill, Claude hands back a pleasant paragraph. With it, Claude hands back a prompt with a subject, an action beat, a camera move, a lighting note, and a sound cue, every time, in the same order.
The consistency is the real payoff. You stop rewording the same request and start comparing results across clips, which is the only way to tell whether a change helped.
Any current Claude model can follow a skill. Claude Sonnet 5 is a sensible default for drafting. For long briefs, say a ten-shot sequence with continuity rules, try Claude Fable 5, or Claude Opus 4.7 when you want it to look at a reference still and describe what it sees. All three are available on PicassoIA if you would rather draft prompts there and paste them into the video page.
Here is a working starting point. Save it as SKILL.md inside a folder named seedance-video. In Claude Code, drop the folder into a skills directory under your project's .claude folder. In the Claude apps, zip the folder and upload it in the skills settings. Either way, connect PicassoIA in your connector settings so the video tools exist.
The file opens with a header fenced by a line of three dashes above and below. Inside it sit two fields:
name: seedance-video
description: Write Seedance 2.5 video prompts and render clips. Use when the user asks for an AI video, a short clip, b-roll, or an animation of an image.
The description does the heavy lifting. Claude matches your request against it, so name the triggers in the words you actually type: "video", "clip", "animate this image".
Below the header, write the rules. This version is short on purpose:
# Seedance video prompts
Write one paragraph in this order:
1. Subject and starting pose
2. Action, in the order it happens
3. Camera move: static, slow push in, tracking, handheld
4. Light and atmosphere
5. Sound: ambient, effects, music, dialogue in "double quotes"
Rules:
- One idea per clip. Split multi-scene briefs into separate clips.
- Photoreal by default unless the user names a style.
- No invented on-screen text. One or two words at most.
- 5 seconds for a single beat, 10 seconds for a full action.
- 720p for finals, 480p for quick drafts.
Workflow:
1. Show the finished prompt and the settings. Wait for a go-ahead.
2. Call the video tool and keep the prediction ID.
3. Check status at the suggested interval until it succeeds or fails.
4. On success, show the video URL. On failure, change one variable and resubmit.
Each line earns its place. The order forces the model to commit to a subject before it reaches for adjectives. The one-idea rule exists because multi-action prompts are a common cause of clips where a character drifts or a prop changes shape midway. And the go-ahead step costs you ten seconds while saving a render.
Here is what the formula does to a rough request. You type: a café in the rain, warm and quiet. The skill returns:
A barista in a green apron slides a ceramic cup across a wooden counter inside a small corner café on a rainy evening. A customer's hand lifts the cup into frame. The camera starts static at counter height, then pushes in slowly over five seconds. Warm tungsten bulbs, rain streaking the window behind her, soft reflections on the wet street outside. Sound: steady rain on glass, the clink of ceramic on wood, a low espresso machine hiss, soft acoustic guitar, no dialogue.
Same idea, five decisions made. The model no longer has to guess the subject, the camera, or the soundtrack.
Add a short checklist the skill runs silently before it calls the video tool:
Because the skill is plain text, you can edit it after every batch. When a clip fails in a new way, add one rule, and the fix applies to every later request.
The skill is the fast lane, but the video page is the best place to test a prompt before you bake it into the file. Seedance 2.5 Lite is the model behind the connector, and per its model page Wonder members run it without a per-clip cost, which makes it the right place to iterate.
| Model | Length | Resolution | Best for |
|---|---|---|---|
| Seedance 2.5 Lite | 5 or 10 seconds | 480p, 720p | Fast iteration, first and last frames |
| Seedance 2.5 | Up to 30 seconds | 480p, 720p | Reference images, lip sync, motion transfer |
Start with Lite. Move to the full model when the brief needs something Lite lacks: a clip longer than 10 seconds, a character locked across shots with reference images, or a voice recording the lips must follow. Seedance 2.0 is still listed if you need to match an earlier project.
| Setting | Options | Advice |
|---|---|---|
| Duration | 5 or 10 seconds | Default is 5. Use 10 only when the action needs room. |
| Resolution | 480p or 720p | 480p renders fastest. 720p is sharper. |
| Aspect ratio | Match input image, 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3 | Ignored when you supply an image, because the clip follows the image. |
| Seed | Any integer | Lock it to repeat a result or to compare edits fairly. |
| Save audio | On or off | On by default. |
Sound is where Seedance separates itself from silent generators: each clip is rendered with synchronized audio rather than arriving as footage that still needs a sound pass. Your skill should treat audio as part of the prompt, not an afterthought.
On the full Seedance 2.5, dialogue goes in double quotes inside the prompt, and the model renders the line with matching mouth movement. Keep the line short enough to fit the clip: roughly one sentence for 5 seconds, two for 10. Add this to the skill: "Spoken lines go in double quotes, one speaker per clip." Test the habit on Lite before you rely on it.
"Sad music" leaves too much to chance. "Slow solo piano, no drums, faint room reverb" gives the model something to build. The same goes for effects.
| Weak cue | Strong cue |
|---|---|
| Stormy atmosphere | Rain on a tin roof, distant thunder |
| Busy street | Car horns, footsteps on wet pavement, a bicycle bell |
| Tense music | Low sustained cello, no melody |
| Happy kitchen | Sizzling pan, a spoon tapping a pot, soft radio chatter |
Ask the skill to list ambient sound, effects, and music as separate short phrases at the end of the prompt.
💡 Silent clips have a place. Turn audio off when you plan to score the piece later, or when a platform will add its own track. The Lite tool exposes a single switch for it.
Text prompts leave composition to the model. An input image takes that decision away from it, which is why image-to-video is the dependable route for brand work, product shots, and any character you need to recognize.
When you supply a picture, the clip opens on it and inherits its aspect ratio. So the prompt should describe motion only. Do not re-describe what is visible; the model can already see it. Add one rule to the skill: "If the user supplies an image, write only what moves, how the camera moves, and what the sound is."
Claude can also make the still for you. Ask it to generate a photoreal first frame with the connector's image tool, approve it, and then pass that image into the video call. Two steps, one conversation.
Both Seedance 2.5 Lite and the full model accept a closing frame. It requires a first frame, and the clip travels from one to the other. This suits product reveals, a door opening onto a known room, or any shot that must end on a specific composition. On the full Seedance 2.5, first and last frame modes also require the aspect ratio set to adaptive, so tell the skill to switch it automatically.
The full model can also take reference images, called out in the prompt as [Image1], [Image2], and so on. Reference images cannot be combined with first or last frame images, so choose one approach per clip.
Clips fail in repeatable ways, which means a skill can prevent most of them.
| Problem | Likely cause | Fix to add to the skill |
|---|---|---|
| Subject drifts or changes shape | Too many actions in one clip | One action per clip |
| Camera ignores the brief | Camera note buried at the end | Put the camera move third, in a plain verb |
| Garbled words on screen | Prompt invented signage | Forbid on-screen text, or allow one quoted word |
| Sound feels generic | Mood words instead of sources | Name each sound source |
| Clip ends mid-action | Action longer than the duration | Cut the action or switch to 10 seconds |
Lock the seed, edit a single sentence, and render again. If the clip improves, keep the change. If it gets worse, you know which sentence to blame. Ask the skill to record the seed with each prompt so you can return to any version.
The platform runs a limited number of predictions at once per account, shared across connections. Ask Claude to call the account tool first to check your plan and parallel limit, then queue clips in small batches instead of twenty at once. A skill rule such as "submit at most three clips at a time, wait, then continue" keeps a long session from stalling.
You now have the whole loop: a skill that writes the prompt, a connector that renders it, and a set of rules that turn failures into edits. The quickest way to test it is to skip the setup and run one prompt by hand.
Open Seedance 2.5 Lite on Picasso IA, paste a five-line prompt in the order above, set 720p and 5 seconds, and watch what comes back. Then generate a first-frame image with the platform's image tools and animate it. If a clip misses, change one line, keep the seed, and run it again. When the prompts stop failing, paste your winning formula into SKILL.md, and Claude will write it every time. For anything longer than 10 seconds, or any shot that needs reference images, step up to Seedance 2.5.
Everything above is one prompt away. Pick a scene you can describe in one breath, give it a camera move and a sound, and see how it looks on screen.
