How to use the Seedance Short Prompt Compressor agent skill

What the skill makes the agent do

Seedance 2.0 is an AI video generator. Long, wordy descriptions often bury the details the model needs. This skill asks the agent to compress the prompt instead of expanding it. A compact prompt still includes the subject, the action and its visible endpoint, one camera move, a physical light source or atmosphere, a sound cue or silence instruction, and key constraints.

The skill removes filler first: generic adjectives, duplicate style labels, obvious background details and secondary actions. It keeps reference tags such as @Image1 and their roles, and it preserves continuity notes. For storyboards, it keeps shots in order and writes each cut in words.

• Output target: about 30-100 English words, or an equivalent Chinese prompt when you ask for Chinese or maximum compression.
• Mixed-language prompts keep only the language mix that clarifies reference roles, dialogue, camera terms or safety constraints.
• The skill refuses requests that a moderation check would block, even when the prompt is short.

Install the skill from the ZIP download

Download the ZIP for the skill from the directory page. The skill is a single file, SKILL.md, from the repository Emily2040/seedance-2.0 under skills/seedance-prompt-short. Unzip it so the skill folder sits in one of the locations below.

You can install it for yourself across projects by unzipping into your personal skills folder, or for one project only by unzipping into that project's .claude/skills/ folder. Claude apps also accept an upload: open Settings, then Capabilities, then Skills, and add the skill there.

• Personal install: place the folder in ~/.claude/skills/
• Project install: place the folder in .claude/skills/ inside the project
• Claude apps: Settings → Capabilities → Skills, then upload the skill

Trigger the skill with natural requests

You do not need a special command. The skill's description tells the agent when to use it, so ordinary requests work. Ask for a compact prompt, a short Chinese prompt, prompt compression, a 30-100 word output, or removal of unnecessary prompt language.

Say what you have and what you want the clip to show. If you already have a long prompt, paste it and ask the agent to shorten it. If you are building a storyboard, say how many shots you want, since the skill keeps shot order and can write one compact line per shot.

• Compact a long idea: 'Make a compact Seedance 2.0 prompt of about 60 words for a rain-soaked street scene with one slow dolly-in.'
• Shorten a draft: 'Compress this Seedance prompt to under 100 words and keep the @Image1 reference.'
• Chinese output: 'Give me a short Chinese Seedance prompt for a close-up of a tea cup, with the sound of rain.'
• Storyboard: 'Turn these three shots into a compact multi-shot Seedance prompt, keeping the shot order.'

What you need to run it

The skill only writes text. It does not generate video. To use its prompts you need access to Seedance 2.0 itself. The skill also depends on referenced files from the same repository: directors-read, moderation-prescreen, shot-table, multishot-grammar, direct-for-the-model and multilingual-community-examples. The SKILL.md file alone links to them, so install the full set of reference files alongside it or the agent may not be able to follow the guidance fully.

• Seedance 2.0 access to run the generated prompts
• The referenced files from the repository, not only SKILL.md
• An agent that loads skills, such as Claude with the skills folder or Settings upload

Safety notes: scripts and network

According to the skill's metadata, it runs no scripts and uses no network. It is instructions and templates that shape the agent's written output. Because it is plain text, you can open SKILL.md and read exactly what it asks the agent to do before you install it.

The skill is released under the MIT licence. Its moderation guidance applies to short prompts as well as long ones, so a compressed prompt that would be blocked at full length is still refused. Compression never removes a lock line, a stated position or a feeling word.

• Runs scripts: no
• Uses network: no
• Licence: MIT
• Do not use it to request content that moderation rules would block

Tips for better compressed prompts

Give the agent the physical details that matter: who or what is in frame, what they do, where the camera moves, and where the light comes from. The skill keeps these and cuts the rest, so vague input produces vague output. If a storyboard cell cannot be filled, the agent is supposed to ask you a question before delivering, which is a useful check.

If the agent drops a detail you need, ask it to restore that one item rather than the whole prompt. Read the one-line note the skill adds when it removed something important, since it tells you what was cut.

• Name reference tags exactly, such as @Image1, so their roles survive compression.
• Ask for one camera move per shot; secondary moves are cut first.
• For longer detail, use a full prompt instead, because compression removes detail by design.