Video Intent Studio

Video Intent Studio

Follow a user-led workflow for video generation. Do not choose the final prompt for the user. Suggest options, keep the current prompt transparent, and ask for confirmation before generation.

Core behavior

• Respond in the user's language. Default to Chinese if the user is writing in Chinese.
• Keep the workflow layered: intent -> ranked video types -> parameter tuning -> prompt confirmation -> generation.
• Present 3 to 5 options at the type-selection step. Keep the option set stable and only reorder by relevance.
• Show the current prompt preview whenever parameters change.
• Let the user revise or exit at every stage.
• Keep a simple state object in memory or scratch space:

state = {
    "user_input": "",
    "selected_type": None,
    "params": {
        "duration": 8,
        "ratio": "16:9",
        "motion": "medium",
        "style": "original",
        "brightness": "normal",
        "subtitle": "off",
        "dream_filter": "off",
    },
    "final_prompt": "",
    "stage": "init",
}

Skill directory and script paths

When this SKILL.md is loaded, resolve the skill directory from the absolute path of this file.

• Suggestion and prompt builder:
  • /scripts/video_agent_backend.py
• Final generator:
  • /scripts/generate_ark_video.py
• Type catalog and defaults:
  • /references/video-types.md
• Usage walkthrough and examples:
  • /references/usage-guide.md

Use absolute paths when running the scripts.

Recommended workflow

1. Capture the user's raw idea.

   • Keep the original wording in state as user_input.
   • Do not rewrite it into a final prompt yet.

2. Rank video types.

   • Run:

python "/scripts/video_agent_backend.py" suggest --input "user idea"

• Present the top 3 to 5 results as numbered options.
• For each option, include:
  • type name
  • one-sentence use case
  • default duration and ratio
  • short reason why it matches

1. After the user chooses a type, build a prompt preview.
   • Run:

python "/scripts/video_agent_backend.py" build ^
  --input "user idea" ^
  --type cinematic-story

• Show:
  • current prompt preview
  • current parameter summary
  • a short numbered list of tunable options

1. If the user adjusts settings, rerun build with explicit parameters.

   • Supported parameters:
     • --duration 5|8|10|12
     • --ratio 9:16|16:9|1:1|4:3
     • --motion light|medium|strong
     • --style realistic|anime|cinematic|original
     • --brightness moody|normal|bright
     • --subtitle off|on
     • --dream-filter off|on
     • --notes "extra user constraint"

2. Before generation, show the final prompt and ask for confirmation.

   • Use a short confirmation question such as:
     • "Final prompt and parameters are ready. Generate now?"

3. After explicit confirmation, generate the video.

   • Run:

python "/scripts/generate_ark_video.py" ^
  --prompt "final prompt text" ^
  --output "C:\path\to\result.mp4"

1. Report success or failure clearly.
   • On success, give the downloaded file path and task id if available.
   • On failure, bucket the issue into one of:
     • API key or auth problem
     • network or polling problem
     • task failed remotely
     • response did not include a downloadable video URL

Important implementation notes

• The bundled generator script intentionally mirrors the existing HTTP + polling pattern already used in this workspace.
• The current API request sends a text prompt payload only. Duration, ratio, motion, style, and other controls are encoded into the prompt text unless you later extend the API payload.
• The generator script reads credentials from environment variables first:
  • ARK_API_KEY
  • VOLCENGINE_ARK_API_KEY
• Optional environment variables:
  • ARK_VIDEO_MODEL
  • ARK_VIDEO_TASKS_URL

Conversation rules

• Do not skip the type-selection step unless the user explicitly says they already know the type.
• Do not ask broad open-ended questions if a numbered choice is possible.
• If the user says "more realistic", "more cinematic", "shorter", "vertical", or similar, treat that as a parameter update and keep moving.
• If the user says "generate", "go", or "就这样生成", show the final prompt once and ask for one explicit confirmation unless they already confirmed in the same message.

When to read references

• Read references/video-types.md when you need the fixed type list, defaults, or category-specific prompt leads.
• Read references/usage-guide.md when you need example conversations, sample commands, or the user-facing tutorial flow.