菜单
Re-transcribe a D&D session's Zoom .m4a recording into a more accurate WebVTT, using the campaign's own proper-noun vocabulary while keeping Zoom's existing speaker/timing boundaries. Runs faster-whisper on the DGX Spark via ~/src/mytools/audio-to-vtt. Invoke as /audio-to-vtt [audio-path] [vtt-path].
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/kostadis/mytools --skill audio-to-vtt命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Propose→review→apply mechanical-residue scrub for session_doc narration prose. A deterministic regex pass surfaces candidate residue (raw numbers on DC/AC/HP/damage/healing/feet/rounds/initiative, out-of-fiction table-speak, real player names as speakers) — never spell names or magic vocabulary, which are not a candidate category at all. The GM confirms every candidate one at a time before anything is rewritten; a deterministic apply step then writes `.scrubbed.md`. Replaces the autonomous `scrub_mechanics.py` LLM pass per CampaignGenerator issue #151 (the spell-stripping incident). Invoke as /scrub [narration-dir-or-file].
Quote-level consistency check on scene_extractions_new/ — flags VTT transcription errors, garbled phrases, and clarity issues in verbatim quote blocks, then proposes corrections for user approval. Invoke as /session-summary-consistency [session-dir].
Run consistency checks at every LLM-pipeline boundary (gm-assist → session-summary → scene extractions → narration) with a human-review gate between stages. Use when the user invokes /staged-consistency [session-dir] and wants the multi-stage check rather than a one-shot. Prevents fix-propagation drift where stale per-scene quotes silently re-inject errors into the next narration run.
基于 SOC 职业分类
正在显示 SKILL.md
| name | audio-to-vtt |
| description | Re-transcribe a D&D session's Zoom .m4a recording into a more accurate WebVTT, using the campaign's own proper-noun vocabulary while keeping Zoom's existing speaker/timing boundaries. Runs faster-whisper on the DGX Spark via ~/src/mytools/audio-to-vtt. Invoke as /audio-to-vtt [audio-path] [vtt-path]. |
| tools | Bash, Glob, Read, AskUserQuestion |
Drives ~/src/mytools/audio-to-vtt/retranscribe.py through its
recommended dry-run -> smoke-test -> confirm -> full-run sequence, so the
user never has to remember the flags or babysit a multi-minute Spark job
without knowing how long it'll take first.
This skill is operational glue only. It does not reimplement any of the
tool's logic (VTT parsing, vocabulary gathering, Spark orchestration,
glossary pass) — read ~/src/mytools/audio-to-vtt/README.md and
CLAUDE.md if you need the design rationale or Spark-side setup/build
details. This skill assumes the Spark side is already set up (see that
README's "Setup" section) — it does not perform first-time Spark
bootstrap.
From $ARGUMENTS:
<audio> <vtt>.summaries/*/ one
level down, if present) for *.m4a files. For each candidate, look for
a sibling *.transcript.vtt (same stem, or the most recently modified
.vtt in the same directory if the stem doesn't match exactly — Zoom's
own export naming isn't always identical between the audio and caption
files). If more than one plausible pair exists, list them and ask which
one via AskUserQuestion rather than guessing.Do not pass --campaign-root unless the files sit outside a campaign
checkout (auto-detection walks up from <vtt> looking for
docs/entity_registry.yaml or notes/vtt_transcription_corrections.md,
and works for the normal <campaign>/summaries/<date>/ layout).
python ~/src/mytools/audio-to-vtt/retranscribe.py <audio> <vtt> --dry-run
Show the user the cue-group count and vocabulary-candidate count/preview from the output. This step never touches the Spark — it's a fast sanity check that the VTT parsed correctly and a campaign root was found.
python ~/src/mytools/audio-to-vtt/retranscribe.py <audio> <vtt> --smoke-test < /dev/null
Redirect stdin from /dev/null (or otherwise ensure no interactive input
is available) — the script's own "Continue with the full session? [y/N]"
prompt is meant for a human at a real terminal, not this skill; it
defaults to "no" on EOF and simply stops after printing the measurement.
That's what you want here: read the printed device_used and RTF, and
the extrapolated full-session estimate, from its output.
Use AskUserQuestion with the real numbers from Phase 3, e.g.:
"Smoke test measured RTF X (device: cuda/cpu). Full session (~N min of audio) is estimated at ~M minutes. Run the full transcription now?"
Options: proceed now / not right now. Do not proceed without an explicit yes — a smoke-test estimate is informative, not a green light on its own, and full runs can take anywhere from a few minutes to hours depending on whether GPU is actually engaging (see CLAUDE.md's hardware notes for why that isn't guaranteed on every box).
If confirmed:
python ~/src/mytools/audio-to-vtt/retranscribe.py <audio> <vtt>
This can run for a while on a long session — prefer run_in_background
if your environment supports it, and report back when it completes rather
than blocking silently. Report the two output paths it prints
(<stem>.retranscribed.vtt and <stem>.retranscribed.cleaned.vtt) and
the glossary replacement count.
Always end by telling the user to run:
/vtt-spell-pass <stem>.retranscribed.cleaned.vtt
so any new unknown proper nouns the re-transcription introduced (or
didn't fully fix) go through that skill's human-confirmation loop. Do not
invoke /vtt-spell-pass automatically — it's a separate, interactive,
per-cluster confirmation flow the user should run deliberately.