How to use the Seedance Continuation Director agent skill

What the seedance-continuation skill makes the agent do

An agent skill is a folder of instructions that an AI agent loads when a task matches the skill's description. This one applies when you are making a multi-clip video with Seedance 2.0 and want to continue, extend, or repair a clip.

The skill makes the agent work from accepted footage. It asks for the real final frame or a description of it, and it refuses to invent the bridge if that information is missing. It also helps choose the continuation type: a seamless continuation, an intentional cut, a bridge between two known states, a repaired tail, or a re-anchored clip when the look has drifted.

• Seamless continuation: same shot, same geography and motion, accepted previous footage as the source.
• Intentional next shot: an editorial cut where story continuity matters but exact frame continuity is not promised.
• Bridge between known states: connects a defined start state and end state, often with first and last frame generation if your tool supports it.
• Repair tail: fixes or regenerates a failed final few seconds before you continue, because continuing from a failed tail carries the error forward.
• Re-anchor after drift: returns to canonical identity and references when character, detail, geography, motion, or audio has degraded.

Install the skill from the ZIP

Download the ZIP for skills/seedance-continuation from this page. The skill is distributed as a single file, SKILL.md, inside the repository skills/seedance-continuation folder.

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 for that project.

Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.

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

Check the requirements before you start

This skill is a parent-linked add-on to the seedance-20 skill, so you need Seedance 2.0 access. You also need an accepted previous clip or its final frame, uploaded to the client you are using.

The skill expects reference files from the same repository: directors-read, continuation-handoff, sequence-project-state, prompt-compiler, reference-transfer-contract, continuity-qc, failure-atlas, and directing-engine. Check that these are available alongside the skill before relying on it.

An optional local helper, scripts/extract_last_frame.py, can pull the final frame from a take. The skill itself does not use the network.

• Seedance 2.0 access
• An accepted clip or its final frame
• Optional: python scripts/extract_last_frame.py <take> to extract the final frame locally
• Reference files from the repository listed above

Trigger the skill with natural requests

You do not need special syntax. Describe what you want in plain language, and mention the accepted footage. The skill's trigger covers requests to continue, extend, make the next part, repair the tail, bridge between known frames, re-anchor drift, or create a successor prompt from accepted footage.

If you attach the clip or final frame and your client can open it, the agent reads what is visible and asks you only about what the image cannot show, such as open motion, camera movement phase, or audio phase. If your client accepts the file but cannot display it, the agent will ask you to describe the ending instead.

• "Continue this clip. Here is the final frame. Make the next part where she reaches the door."
• "The last two seconds of take 4 failed. Repair the tail before we go on."
• "Bridge from this final frame to the car interior shot I already have."
• "The character's face has drifted in the last clip. Re-anchor it from our canonical references."
• "Write a successor prompt for the next clip from this accepted footage."

What the agent will ask you for

Before writing a continuation prompt, the agent needs a set of inputs. Many are project bookkeeping, such as a project ID, the current clip ID, and the parent clip ID. The most important input is the observed end state of the accepted clip. The agent will not guess it.

If you do not have the actual footage, the agent will tell you it has the story plan but not the real ending, and ask you to upload the clip or describe exactly what is visible at the end. It should not invent the continuation state.

• Project ID, current clip ID, and parent clip ID of an accepted clip
• Scene ID, and whether the next clip stays in the scene or crosses a scene boundary
• The next clip's intent: what the viewer should feel or notice
• The accepted final frame or a precise description of the end state
• Continuity locks and the reference registry

Read the output and confirm the intent

The skill returns a structured answer: the continuation type, the source evidence it used, the observed end state, the next clip contract, an intent line such as "this clip exists so the viewer feels X", continuity locks, beats to exclude, and the final Seedance prompt for the current clip only.

Confirm the intent line before you generate. The skill asks for this check so you do not spend generation on the wrong clip.

• Continuation type
• Observed end state
• Continuity locks and allowed changes
• Intent echo, confirmed before generation
• Final prompt for the current clip only

Safety notes on scripts and network

According to the listing, this skill runs scripts and does not use the network. The script it references, scripts/extract_last_frame.py, is an optional local helper for pulling a final frame from a take.

Read any script before running it, and run it on your own files. Because the listing says the skill runs scripts, treat it the way you would treat any code you install.

• Runs scripts: yes (optional local helper)
• Uses network: no
• Review scripts/extract_last_frame.py before running it on your footage

Tips for better continuations

Give the agent the real ending, not the plan. The skill treats accepted footage as the only truth, so a precise description of the last visible moment produces a better prompt than a summary of the story outline.

Repair a failed tail before continuing. Continuing from a broken final few seconds tends to amplify the error.

Expect a cut at scene boundaries. Crossing into a new scene is treated as an intentional next shot from canonical references, not a seamless continuation.

Re-anchor early. Visible drift is a reason to return to canonical references before the chain grows longer.

• Upload the final frame or describe the exact visible end.
• Repair failed tails first.
• Plan a cut at scene changes.
• Re-anchor as soon as drift is visible.