UNFRAMED Local Story

Use this Skill only on a local folder the user explicitly provides.

Safety contract

• Never move, rename, delete, overwrite, or upload source media.
• Treat identity, emotion, relationships, and the meaning of absence as hypotheses, not facts.
• Include source filename and timestamp or image index for every visual claim.
• Ask before any external publication or cloud-generation step.

Workflow

1. Confirm the input folder and the output folder.
2. Start or check the local service described in references/local-api.md. From this Skill directory on macOS, run scripts/local-service/scripts/start_local_service.sh in a dedicated terminal and keep it open.
3. Create a read-only media manifest.
4. Request local feature extraction and similarity analysis from localhost.
5. Produce three story hypotheses, each with supporting and contradicting evidence.
6. After the user chooses one hypothesis, generate:
   • a 60–90 second beat sheet;
   • a shot list with source references;
   • narration and subtitles;
   • a missing-shot / generation list;
   • copyright, portrait, voice, music, font, and AI-label checks.
7. Save machine-readable JSON and a concise Markdown report.

Output rules

• Separate observations, inferences, and creative_proposals.
• Use confidence labels high, medium, or low; do not invent percentages.
• Prefer a small, filmable human event over spectacle.
• If the local AI backend is unavailable, stop with a setup message; do not silently replace it with cloud AI.

Validation

For the competition demo, run the supplied sample and retain:

• the Qoder prompt;
• the Skill invocation indication;
• localhost health and analysis output;
• the generated JSON/Markdown pack;
• one screen recording showing the full path from instruction to result.

Run the bundled tests before publishing:

/usr/bin/python3 -m unittest discover -s scripts/local-service/tests -v

This MVP uses Apple Vision locally for feature-print similarity and face rectangle detection. It does not identify people or infer relationships. For videos, it analyzes one middle representative frame. State these limitations in every public description.