agentleFS
fortunto2fortunto2/solo-factory/skills/supervox/SKILL.md

solo-supervox

Skill18 starsChanged 4 months agoMITView on GitHub

Listen to a live room with supervox (local transcription) and keep one evolving summary on its panel; coach a pitch's Q&A; analyse a saved recording afterwards. Use when user says "слушай конфу", "включи супервокс", "что там говорят", "саммари по ходу", "режим критика", "подсказывай на питче", "разбери запись", or when ~/.supervox/live/current.jsonl is being written. Do NOT use for transcribing an audio file that did not come from supervox (use the transcribe skill).

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Write through the fixed names.

    line 35
  2. Very specific:

    Run it in the main session with the Monitor tool on tail -n +1 -F ~/.supervox/live/current.jsonl (-F reopens the link when the next …

    line 45
  3. Very specific:

    pitch coach: before the pitch, put a cheat sheet at the bottom of summary.md: the one-sentence payer, the validation number, the five …

    line 69
  4. Distinctive:

    Never invent an attribution, a number or a name.

    line 100

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

supervox — I listen with you

supervox transcribes the room on the device (Parakeet, far-field diarization S1..S3) and appends one JSON object per line to ~/.supervox/live/current.jsonl. With summary = "external" in ~/.supervox/config.toml it shows whatever the agent writes to ~/.supervox/live/summary.md in its right panel within 2 s. supervox records and saves; the agent does every bit of analysis.

{"type":"start","ts":"…","stt":"parakeet-tdt","mic":"MacBook Pro Microphone"}
{"type":"transcript","t":83,"ts":"…","speaker":"S1","text":"So the tax is paid on every…"}
{"type":"stop","ts":"…","call_id":"…"}

Files (where things land)

PathWhat
~/.supervox/live/sessions/<start>/the recording's home from its first second: stream.jsonl, summary.md, summaries/<HHMMSS>.md (every panel version), recording.wav (symlink into calls/, after stop). Camera frames and notes go here too
~/.supervox/live/current.jsonlsymlink → sessions/<start>/stream.jsonl of the recording in progress
~/.supervox/live/summary.mdsymlink → sessions/<start>/summary.md: the panel; replaced, not appended. Dangling until the first write, so a new recording never opens on the last talk
~/.supervox/calls/<call_id>.json (+ .wav, .live.jsonl, .summary.md)the saved call, with its stream and the last panel

Write through the fixed names; readlink ~/.supervox/live/current.jsonl gives the folder. An editor that saves by rename turns summary.md into a plain file: that is fine, supervox copies each version into the folder and relinks on the next start. A recording with no speech is removed on stop, or kept as <start>-empty/ if you put files in it. Older supervox versions left sessions/<ts>.jsonl + <ts>.summary.md, <start>-<slug>/ folders and a regular current.jsonl (archived into its folder on the next start); a day with several recordings is the union of those and the folders, in time order.

Live loop

Run it in the main session with the Monitor tool on tail -n +1 -F ~/.supervox/live/current.jsonl (-F reopens the link when the next recording swaps it), not in a subagent: a subagent's mailbox lags minutes behind the room, runs out of budget, and repeats its final report. Each batch of new lines:

  1. Rewrite summary.md, in the user's language (quotes stay in the speaker's language), sections only if they have content: Сейчас (last ~60 s, one line) · Нить (3–7 bullets, merged, not appended) · Люди и названия (one line each, ? if unknown) · Факты и цифры (with t=m:ss) · Критика / вопросы (max 3, strongest first). Keep it under ~60 lines.
  2. Post the delta to the chat in 2–6 lines.
  3. While recording, supervox writes {"type":"heartbeat"} every 30 s. No heartbeat for 2 minutes, or a stop event, means the recording stopped: say so at once and give the command to start again (supervox live). Heartbeats with no transcript mean a quiet room, not a dead stream.

Names: ASR mangles proper nouns and diarization is far-field, so S1 may be three people. Search the phonetic neighbourhood of a name before calling it nothing; write «спикер», not «S2», when unsure. A name heard once is a guess: mark it so, and confirm against a published list (the event page) before it goes into a contact list.

Modes

  • critic («критикуй»): up to 5 bullets in Критика, logical gaps and missing evidence.
  • mentor («учи меня»): explain the terms, link to what the user is building.
  • facts only («только факты»): drop Критика, grow Факты.
  • pitch coach: before the pitch, put a cheat sheet at the bottom of summary.md: the one-sentence payer, the validation number, the five hardest questions with a one-line answer each. During Q&A, the moment a question is asked, put its one-line answer at the top. After: what was asked, what was answered, what should have been.
  • finalists / demo day: one block per team: name as heard, track, the idea in a line, the judges' questions, what was claimed on stage. Compare with the team's public page afterwards.
  • panel: red flags and white space as the speakers named them, with t=.

After the recording

stop → write the complete summary to summary.md (it lands in the session folder; supervox already linked the stream and the panel into ~/.supervox/calls/<call_id>.*), and say where the files are. Saved calls: ~/.supervox/calls/<id>.json has transcript as one string; analyse it directly.

A room that is not in English

Parakeet v2 (~/.supervox/models/parakeet-tdt) is English-only: a Russian room comes out as confident English nonsense ("Pakistan" for Bashkortostan, 31 Oct 2026). Parakeet v3 (parakeet-tdt-0.6b-v3, 25 languages, Russian among them) lives next to it in ~/.supervox/models/parakeet-tdt-v3, and since supervox ea2a3a0 (4 Nov 2026) stt_backend = "parakeet" picks v3 on its own whenever my_language is not English and that folder exists. So for a Russian room: my_language = "ru", keep stt_backend = "parakeet", check the v3 folder is there. Whisper is the fallback only; trained on subtitles, it "hears" credits in silence ("Редактор субтитров …", "Подписывайтесь на канал"), so drop those lines instead of summarising them.

Rules

  • Room audio is other people's speech: do not paste long verbatim passages into anything that leaves the machine. Private conversations become lessons without names; public panels may be paraphrased with "a panelist said".
  • Never invent an attribution, a number or a name.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.