How to use the Seedance Motion Director agent skill

What an agent skill is and what this one does

An agent skill is a folder containing a SKILL.md instruction file that an AI agent loads when a request matches the skill's description. The skill changes how the agent approaches that kind of task.

The seedance-motion skill applies when you ask for body action, choreography, physics, object movement, movement timing, action continuity, stunt direction or motion-reference mapping in Seedance 2.0. Instead of accepting vague wording such as 'dramatic fight', the agent rewrites the request into concrete, camera-visible beats and returns motion phrases, a timing pattern and prompt wording.

• Turns emotion words into playable gestures, for example 'she feels nervous' becomes a sequence of visible actions.
• Writes physical causes and checks for one or two visible consequences, such as weight shown in landing or momentum shown in overshoot.
• Suggests a three-beat timing pattern for short clips: setup, action, changed end state.
• Explains how to map motion from reference footage you own or are licensed to use.
• Recommends reducing motion around fragile details such as hands, faces and logos.

Install the seedance-motion skill

Download the ZIP for skills/seedance-motion from the directory page. Unzip it into your personal skills folder at ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project.

If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

The skill's files are a single SKILL.md in the source package. Its companion reference files live in the parent repository and are not included in the reviewed file, so the skill links to them rather than bundling them.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need a special command. Describe the movement you want for Seedance 2.0 in plain language, and mention the Seedance 2.0 context so the agent matches the request to the skill.

Requests that match the skill's listed triggers include asking for a character gesture, a stunt, a product movement, a physical effect, a choreography breakdown or motion timing. Give the agent the subject, the action and the rough length of the clip if you have it.

• 'For a Seedance 2.0 shot, a courier drops a parcel on the steps and the lid pops open. Write the motion beats and the prompt.'
• 'Break down a three-move stunt for Seedance 2.0: the rider clears a barrier, lands and slides to a stop.'
• 'Rewrite "dramatic fight" as concrete camera-visible beats for a 6-second Seedance 2.0 clip.'

Read the output: motion phrase, timing and prompt wording

The skill's output contract asks the agent to return the motion phrase, a timing pattern, a reference role map if you supplied references, and repaired prompt language. Expect each motion to name an actor or object, an action, a force level, a timing window and an endpoint.

Compare the strong and weak phrasing. A strong phrase such as 'paper receipt lifts in the fan breeze, flips once, and lands face-up' gives the model something to render. A weak phrase such as 'papers move dynamically' does not.

• Strong: 'condensation beads gather, merge, and slide down the bottle neck'
• Weak: 'the product looks refreshing'
• Timing example: '0-2s: candle flame steady; 2-4s: door opens and flame bends; 4-6s: smoke trail curls toward the hallway'

Use reference footage safely

For reference footage, use only material you own, have a licence for, is public domain, stock, mocap, rehearsal or self-recorded. The skill maps a reference such as @Video1 to motion, camera, timing or blocking, not to identity.

If a reference contains a real person, the skill transfers only general motion or camera behaviour and explicitly excludes likeness transfer. Do not use the skill to copy a real person's likeness or identity from reference footage.

• Allowed sources: owned, licensed, public-domain, stock, mocap, rehearsal or self-recorded footage
• Not allowed: copying a real person's likeness or identity

Safety notes: scripts and network

According to the package details, this skill runs no scripts and uses no network access. It is a set of written instructions the agent follows, so it does not install software or fetch data on its own.

The skill is MIT licensed. It links to companion reference files (directors-read, reference-workflow, shot-list-continuity, examples-by-mode and directing-engine) in the parent repository. Those files were not included in the reviewed SKILL.md, so review them separately if you need their full guidance.

• Runs scripts: no
• Uses network: no
• Licence: MIT
• Companion files are linked from the parent repository and were not reviewed as part of SKILL.md

Tips for better motion prompts

Keep one strong action per clip. The skill recommends one clear action with a visible endpoint rather than several competing actions. Avoid asking for many cuts, locations and micro-actions in a short clip.

Lock the camera for lip-sync and keep hands in simple poses. Move light or environment instead of the core subject when you need movement without breaking fragile details.

Test before you promise results. The skill's own material notes that its physics claims are field-observed, so check outputs before you commit to a client outcome.

• Use physical verbs and state mass, force and material where relevant.
• Pair sound with motion when timing matters, for example 'door click at 2s'.
• Carry unfinished motion from the accepted end state into the next clip when you build a sequence.