media-transcription-report
Installation
SKILL.md
Media Transcription Report
Use this skill when local audio/video needs to become a clean transcript, or when transcripts and source notes need to become a neutral report in Markdown, TeX, and PDF.
Core Rules
- Preserve original media. Copy external files into the project when needed, but do not delete or overwrite originals.
- Prefer the user's requested model and language. Default to Whisper/WhisperX
large-v2and Chinese (zh) when requested; fall back tomediumandint8compute if GPU memory fails. - Treat speaker labels as optional. If the user asks for speaker info, use WhisperX diarization when credentials and models are available; otherwise produce the transcript immediately and record the speaker-info blocker in the work notes.
- Write public reports as independent third-person documents. Do not mention the chat, recording, transcript, prompt, user request, AI generation, or private working context unless the user explicitly asks for that disclosure.
- Keep raw transcripts, private notes, TeX, figures, and PDFs separated so public artifacts can be shared without leaking source material.
- Validate with actual files: transcript exists, timestamps render, TeX compiles, PDF opens, and generated figures or diagrams do not overlap.
Recommended Layout
Use the project layout when it already exists. Otherwise create a compact structure: