Menu
Processes offline audio recordings from an external mic into the daily log pipeline.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Baseado na classificação ocupacional SOC
Select, score, and source-remap standalone vertical short-form clips from a finished long-form video, creator-edited timeline, transcript, and original camera/audio/screen-recording sources. Use when repurposing a course, podcast, interview, tutorial, livestream, or horizontal A-roll edit into Instagram Reels, TikTok videos, or YouTube Shorts; when the user wants a ranked candidate list before editing; or when an approved long-form cut must be traced back to higher-quality untouched media for vertical reframing.
Direct and execute the visual pass after a short- or long-form talking-head video's A-roll and spoken story are locked. Use when a creator has a versioned A-cut export and wants Codex to apply a creator visual-style profile and produce A-roll punch-ins or reframes, proof assets, B-roll, captions, sound effects, transitions, transcript-anchored full-screen takeovers, diagrams, and Remotion or Manim animation while preserving the locked story. Supports long-form course openings, live Remotion Studio review, full-composite renders, transparent overlay tracks, individual editor assets, and native editor integration.
Umbrella skill for private export and import pipelines. Use this when the task spans multiple data sources or when no more specific export skill applies.
Modular tooling for journal ingestion, prep, and derived-context generation. Use when work involves reusable journal processing logic.
Run IgnitionBot trend sweeps and return scored briefs across configured sources.
Measure the current virality, velocity, and editorial usefulness of a news story, product launch, model release, funding event, or trend. Use when the user asks for a heat score, asks how hot/viral/trending something is, wants to compare competing news waves, decide whether a topic is worth covering now, distinguish mass attention from niche audience fit, or audit why a radar missed a high-signal canonical announcement.
| name | transcribe-offline-meeting |
| description | Processes offline audio recordings from an external mic into the daily log pipeline. |
| metadata | {"version":1,"tags":["audio","journal","transcription"]} |
This skill defines the process for an agent to take a raw .wav or .mp3 file
(typically recorded offline via a DJI Mic 2 or Dictaphone), transcribe it using
the native ARM64 MLX Whisper path, and then use the AI's natural language
understanding to provide a speaker-labeled summary in the LifeRepo
DAILY_SUMMARY_WORKFLOW.
The core transcription utility is located at:
georgeskills/scripts/transcription/mlx_transcriber.py
~/.config/georgeskills/mlx-whisper.env.Use the MLX path for transcription. Do not choose WhisperX,
whisper-ctranslate2, or generic CPU Whisper as a normal fallback for this
workflow. If MLX fails, fix the MLX environment first: native ARM64 venv,
working mlx / mlx_whisper imports, and a valid Hugging Face cache.
georgeskills/scripts/transcription/run_mlx_transcriber.sh /path/to/file.wav --outdir /georgerepo/journal/audio/transcripts/
Current defaults for the MLX path:
mlx-community/whisper-large-v3-turbomlx-community/whisper-large-v3-mlx-4bitHF_HOME to a
stable local cache before running the MLX script.Bootstrap or repair the runtime with
georgeskills/scripts/transcription/bootstrap_mlx_runtime.py. Put the runtime
outside disposable repo artifacts and put the model cache on durable storage.
Do not use an x86_64 uv-managed Python on Apple Silicon.
Do not use the old CPU transcriber to work around a local setup issue. Treat MLX import/cache/model failures as environment blockers and record the blocker in the daily summary or command notes.
Storage rule:
Workspace/dji-audio/ raw-only.<private-repo>/journal/audio/transcripts/YYYY/MM/.<private-repo>/journal/audio/transcripts/ and auto-redirects the legacy dji-audio/transcripts/ and journal/inbox/ targets back into the canonical journal audio tree.After the script generates the raw transcript:
## Conversation Milestones header using the personality guidelines in georgerepo/journal/PRIVATE-journal.md.Summary.md.