World Cup Video Generator

Intent

Turn a user's natural-language request into a finished Chinese match analysis MP4. The user should not need to know commands, package directories, pnpm, or stage names.

Use scripts/worldcup-video.mjs as the only command entrypoint. The skill carries its runtime under assets/engine.

Natural-Language Workflow

When the user asks for a match analysis video:

1. Extract the matchup.
   - Translate common Chinese country/team names to English for CLI arguments.
   - Preserve the user's wording in --note when it adds analysis requirements.
2. Resolve relative dates.
   - "today" means the agent's current local date.
   - "tomorrow" and "tonight" should become an ISO date in --kickoff YYYY-MM-DD when useful.
3. Infer defaults.
   - Use --competition "World Cup" unless the user names another competition.
   - Use no --stage unless the user names a round such as quarter-final, semi-final, final, group stage, or round of 16.
4. Build a stable run id such as spain-belgium-20260710.
5. Run doctor, then generate with --render-video --force.
6. Verify the MP4 exists and report the absolute path.

Ask a follow-up only when one of the two teams is missing, the user asks for a non-World-Cup competition that cannot be inferred, or provider auth is unavailable.

Example

User says:

Generate today's Spain vs Belgium match analysis video.

Run from this skill directory:

node scripts/worldcup-video.mjs doctor
node scripts/worldcup-video.mjs generate --id spain-belgium-<YYYYMMDD> --home Spain --away Belgium --competition "World Cup" --kickoff <YYYY-MM-DD> --note "Generate a Chinese pre-match analysis video for today's Spain vs Belgium match." --render-video --force

If the user says the same thing in Chinese, do the same translation and execution. Do not ask the user to provide command-line flags.

Command Entrypoint

Use these commands internally:

node scripts/worldcup-video.mjs doctor
node scripts/worldcup-video.mjs generate --id <run-id> --home <team> --away <team> --competition "World Cup" --render-video --force
node scripts/worldcup-video.mjs render image --packageDir assets/engine/outputs/runs/<run-id>/video --scene 3 --force
node scripts/worldcup-video.mjs render audio --packageDir assets/engine/outputs/runs/<run-id>/video --force
node scripts/worldcup-video.mjs render avatar --packageDir assets/engine/outputs/runs/<run-id>/video --force
node scripts/worldcup-video.mjs render segment --packageDir assets/engine/outputs/runs/<run-id>/video --force
node scripts/worldcup-video.mjs render assemble --packageDir assets/engine/outputs/runs/<run-id>/video --force

The final video is written to:

assets/engine/outputs/runs/<run-id>/video/<run-id>.mp4

If the user wants outputs somewhere else, pass --output-dir <path>.

Request Parsing

Read references/request-parsing.md for:

• Chinese-to-English team names.
• Relative date handling.
• Run id conventions.
• Final response style.

Failure Handling

If the complete generate command fails:

1. Identify the failed stage from the error.
2. Read references/troubleshooting.md.
3. Rerun only the failed stage when possible.
4. Continue to final assembly after the failed stage succeeds.

Do not restart the entire full pipeline unless search/report/storyboard inputs need to change.

Model Defaults

Keep long reasoning models for report writing only. The engine defaults storyboard JSON and image prompt optimization to qwen-plus because those stages must be fast and schema-stable.

Use these environment variables only when needed:

PI_MODEL_ID                 # report model
PI_STORYBOARD_MODEL_ID      # storyboard JSON model, default qwen-plus
PI_IMAGE_PROMPT_MODEL_ID    # image prompt optimizer, default qwen-plus

Safety Rules

• Do not print provider auth values or .env contents.
• Do not invent unsupported injuries, lineups, odds, match times, venues, or weather.
• Treat odds, probabilities, and score predictions as uncertain pre-match analysis, never betting advice.
• Do not promise outcomes, returns, accuracy, guaranteed wins, or betting actions.

References

• Read references/request-parsing.md for natural-language request handling.
• Read references/examples.md when adapting a user sentence into a command.
• Read references/env.md when provider auth or settings are missing.
• Read references/output-contract.md when verifying or handing off artifacts.
• Read references/troubleshooting.md when a stage fails.
• Read references/engine.md when copying, updating, or rebuilding the portable engine.