Local Creator Assistant

Convert local media into a structured content pack without uploading the source file.

Workflow

1. Confirm the input is a local audio or video path.

2. Choose an output directory. Default to <media-parent>/<media-stem>-content-pack.

3. Invoke only the fixed entry point:

   scripts\run.ps1 "C:\path\to\video.mp4" --output "C:\path\to\content-pack" --language zh
   

4. Read manifest.json after completion and report the produced artifacts.

5. Preserve the source file. Never upload it or replace it.

For macOS/Linux development, use the equivalent test entry:

python3 scripts/client.py "/path/to/video.mp4" --output "/path/to/content-pack" --language zh

Options

| Intent | Option | | --- | --- | | Let Whisper detect language | --language auto | | Force Chinese transcription | --language zh | | Select Intel device | --device NPU, --device GPU, or --device CPU | | Use an exported local model | --model-path C:\models\whisper-small-int8-ov | | Run a deterministic test | --backend fixture --fixture tests\fixtures\sample-transcript.json | | Skip the resident service | --no-server |

Outputs

• transcript.md: readable transcript with timestamps.
• transcript.json: machine-readable transcript and segments.
• subtitles.srt: subtitle track.
• chapters.md: extractive chapter list.
• manifest.json: inputs, runtime metadata, output paths, and timings.

Treat transcript.json as the contract for downstream Agents. Keep its schema_version field when extending it.

Reliability Rules

• Use the local OpenVINO backend by default; never silently fall back to a cloud API.
• Fail with a clear message when the model, FFmpeg, or local runtime is missing.
• Write outputs atomically so interrupted runs do not leave corrupted artifacts.
• Return non-zero exit codes for invalid input, initialization failure, or inference failure.
• Keep UTF-8 encoding for commands, logs, JSON, Markdown, and subtitles.
• Reuse the resident model service for repeated calls.

Boundaries

• Do not claim speaker diarization, OCR, abstractive rewriting, or highlight scoring in this version.
• Do not invent words when audio is unclear; preserve the model result.
• Do not send local media or transcripts to external services.

Read references/deployment.md when installing OpenVINO, exporting a model, selecting CPU/GPU/NPU, or preparing the Intel AI PC demonstration.