agentleFS
prajwal-svmprajwal-svm/product-promo-harness/skills/promo-voice/SKILL.md

promo-voice

Skill1 starsChanged 45 days agoMITView on GitHub

Direct and generate voice-over for a product promo harness run with `promo voice`, using Supertonic (local, free, no key) or Gemini 3.8 TTS (cloud, style-directed acting): choosing the provider, auditions, pronunciation respellings, ASR aliases, verify false for interjections, retakes verified by speech recognition, and fixing CHECK lines. Use when creating, directing, auditioning or fixing narration or voice-over for a promo or explainer video.

  • Reads credentials

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Gemini: when a line must be acted (a whisper, a sigh, a smile on one word, a mock-weary aside), for a language Supertonic lacks, or when …

    line 22
  2. Distinctive:

    Ask the user how any brand word is said before the first take.

    line 45
  3. Distinctive:

    When the audio is right but ASR writes the word differently, list the ways it writes it.

    line 49

≈ “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

promo-voice

promo voice makes, checks and stores every VO take; this skill is how to choose and direct the voice. Measured facts about both providers are in the harness's docs/api-notes.md.

Pick the provider

Supertonic ("provider": "supertonic")Gemini 3.8 TTS ("provider": "gemini")
Runson this Mac, supertonic CLI 1.1.2Google's API
Key and costno key, free (usd: 0 rows in costs.jsonl)GEMINI_API_KEY; cents a film
VoicesF1-F5, M1-M5; F2 is a friendly defaultKore (default), Aoede, Leda, Autonoe, Callirrhoe
Directionnone: no style prompt, inline tags are stripped, punctuation is the only pause controla style string per film, chapter and line
Languagesen, ko, es, pt, frdetected automatically, 130+
  • Supertonic F2 is a good default. It needs no key and takes about a second a line.
  • Gemini when a line must be acted (a whisper, a sigh, a smile on one word, a mock-weary aside), for a language Supertonic lacks, or when the user asks for it.

The pack's default provider is still Gemini, so set it in each run's shotlist.json; unset fields fall back to the pack's voice.json. promo new writes Gemini's voice ("name": "Kore") into the scaffold: change the name too when you switch, or Supertonic is asked for a voice it does not have.

"voice": {"provider": "supertonic", "name": "F2", "steps": 12, "speed": 1.25}
"voice": {"provider": "gemini", "name": "Kore", "style": "warm, confident, conversational, lightly upbeat"}

Supertonic's steps, speed (0.7-2.0) and silence_s are its only controls: steps 12 at speed 1.25 is about 167 wpm, and the pack checks pace against 170. Giving them with "provider": "gemini" is a shot list error.

Pronunciation (pronounce)

Respellings change only what the voice is sent; captions, WER and word anchors keep the real spelling:

"pronounce": {"Kubernetes": "koo-ber-NET-eez", "LaTeX": "Lay-tek", "BibTeX": "Bib-tek", "repo": "ree-poh"}

Ask the user how any brand word is said before the first take.

What speech recognition mishears (asr_aliases)

When the audio is right but ASR writes the word differently, list the ways it writes it; scoring then reads them as the script's word:

"asr_aliases": {"Kubernetes": ["cooper netties"], "Typst": ["tips", "tipst"]}

List only mishearings of correct audio. A garbled name is a wrong take, not an alias. US and UK spellings are already treated as equal.

Interjections ("verify": false)

A pure interjection such as "Psst." or "Hm." cannot be scored by ASR. Put "verify": false on that line: it skips the WER check (pace, clipping and silence are still checked), promo lint-script warns, and promo voice lists it under "Not verified". Tell the user to listen to those lines. Prefer a speakable line where you can.

Commands

  • promo voice audition "<a real line>" --voices F1,F2,M1 (Supertonic) or promo voice audition "<a real line>" --voices Kore,Aoede,Leda --style "..." (Gemini). Files land in audio/auditions/; the user listens and picks; set voice.name.
  • promo voice --all, or promo voice --line <id> (repeatable). Each take is trimmed, normalised to -16 LUFS, transcribed (Parakeet for English, Whisper otherwise), aligned to the script's words and retaken up to 3 times until WER is 0.1 or less and the pace is within 30% of the pack's. Gemini retakes add "clear diction", then "slightly slower, clear diction" to the style; Supertonic retakes change the sampler (steps +4 and speed x0.97, then steps +8 and speed x1.03).
  • A line marked CHECK: fix the text, pronounce or asr_aliases, then promo voice --line <id>. Rescoring cached audio is free: promo voice --all --force. Supertonic draws differ every time and the cache pins the first one, so for a fresh draw change steps or speed.
  • Then promo timing: real takes replace the estimates.

Directing Gemini

  • Models: gemini-3.8-flash-tts for finals, gemini-3.8-flash-lite-tts for auditions and drafts.
  • Cost: $0.50 per 1M text tokens in; $9.00 (flash) or $6.00 (lite) per 1M audio tokens out; 25 audio tokens a second. Prices double on 2027-01-01. costs.jsonl logs every take.
  • The text is read verbatim: "Say cheerfully: hello" speaks the instruction. Direction goes only in the style strings (pack voice.json, shotlist.json voice style, each line's style), sent as speech_metadata.style.
  • Style vocabulary, most important first: emotion, pace, energy, intimacy, smile. "warm, confident, conversational, lightly upbeat; brisk within the line" (ad-sparse); "friendly, quick, playful; a smile in the voice; lean on the key word" (ad-dense); "calm, curious, precise; unhurried; lets important words land" (professor). Keep one style per chapter so takes do not drift.
  • Inline tags inside the text: <short pause>, <long pause>, <breath>, <sigh>, <laugh>. Sparingly; never sound effects. Supertonic would read them aloud, so the harness strips them for it.

Pitfalls

  • Numbers and currency are normalised before WER ("$10,000" = "ten thousand dollars").
  • Short Supertonic lines sometimes come back silent; the harness redraws them, and a line silent in every take is reported as CHECK, never stored as a take.
  • You cannot hear the voice. Never say it "sounds good": say it was verified by speech recognition and loudness.

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.