How to use the Seedance Filter Helper agent skill

What is an agent skill and what does seedance-filter make the agent do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent loads when a task matches the skill's description. This one is named seedance-filter and is licensed MIT.

When it applies, the agent first identifies the requested scene, its must-haves and whatever rejection information is available. It separates the provider's stated reason from the user's report and from any unverified hypothesis, and it leaves the cause unknown when the evidence does not establish it. It asks for missing context only when that context materially affects the boundary or the requested output.

• Separates the provider's stated reason, the user's report and unverified guesses
• Assesses the underlying content before offering any clarification
• Offers a faithful clarification of ambiguous wording, or a clearly labeled alternative
• Explains any material change it makes and any remaining boundary
• Does not promise that a revised prompt will be accepted

Install the seedance-filter skill

Download the ZIP of this skill from our page and unzip it. You can install it in one of two ways, depending on how you use Claude.

For a personal setup, unzip the skill into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

The source material does not describe further setup steps, so check that the folder contains SKILL.md after unzipping.

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

Requirements before you start

The skill references a parent skill named seedance-20, and it refers to reference files that are not included in the provided material. You need Seedance 2.0 access to use it for generation work. The skill does not run scripts and does not use the network.

Because the skill's own file is the only one listed, expect the agent to be limited to the guidance in SKILL.md when those referenced files are missing.

• Seedance 2.0 access is required
• Parent skill seedance-20 is referenced by the skill
• Runs scripts: no
• Uses network: no

How to trigger the skill with natural requests

You do not need a special command. The skill is meant to load when your request matches its description: a Seedance 2.0 prompt was blocked or rejected, moderation may be causing a problem, or you want a content-boundary review or a safer alternative. Describe what happened in plain words and include any error text you received.

Give the agent the actual prompt, the exact rejection message if you have one, and what you were trying to make. The more accurate the report, the better the agent can separate facts from guesses.

• "My Seedance 2.0 prompt was rejected. Here is the message and the prompt. Can you tell me whether this looks like moderation or a generation defect?"
• "Review this blocked Seedance 2.0 prompt and suggest a clearer wording for a product demo video."
• "I think moderation caused this failure. Assess the request before suggesting changes."

Examples of useful requests

Good requests state the benign goal and the observed result. The skill is built to assess the request first, so a clear description of the scene and the rejection gives the agent what it needs.

If the agent proposes a changed scene rather than a wording fix, it should describe that change and label it as an alternative. Treat copyable prompt text and the agent's notes as separate things.

• Ask for a second opinion on whether a block was caused by moderation or by a output defect
• Ask for a clearer wording of an ambiguous phrase that has a known benign meaning
• Ask for a labeled alternative when a scene element is the likely problem

Safety notes: what the skill refuses

The skill clarifies benign requests. It does not disguise prohibited content, help bypass or trick moderation, or help evade safety systems. It refuses evasion and hidden-word tactics, and it states that misuse could violate provider terms.

It also does not assume facts it cannot verify. Requests involving real people's likeness, minors, or consent without proper authorization are not something the skill will assume away. A text revision does not authorize an upload, a retry or spending, so follow your own budget and retake practice. Face-limit or portrait-verification workarounds are not offered; the skill points to official sanctioned flows instead.

• Refuses to disguise prohibited content or evade safety systems
• Does not treat adding words like "prop", "original" or "authorized" as establishing those facts
• Does not guarantee acceptance of a revised prompt

Tips for better results

Keep your original prompt and the exact rejection text together, so the agent can judge both. Say what you want the final video to show in ordinary words. If you already have a budget or retry limit, state it, because the skill follows the user's budget.

Expect honest uncertainty. When the evidence does not show a cause, the agent should say so rather than guess.

• Paste the exact prompt and the exact error text
• State the benign goal in one or two sentences
• Mention any retry or spending limits you want respected
• Read the notes separately from any copyable text