Narrate

Skill AlexAnsart/demo-studio/skills/narrate

Agent skills that turn any web app into a polished, narrated demo video — Playwright capture, smart zooms, ElevenLabs voiceover, Whisper sync and burned captions.

Install

npx -y skills add AlexAnsart/demo-studio --skill narrate

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

11 days oldThe repository was created 11 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Generates a narration voiceover .mp3 for a demo video script using the ElevenLabs API — configurable voice/model/stability, deterministic output filename. Use when the user asks to add a voiceover, narration, or spoken audio to a demo/tutorial video, or names ElevenLabs explicitly.

SKILL.md

3.3 KB, as published. Nobody here has run it

Narrate — generate a voiceover with ElevenLabs

Turns a plain-text script into an .mp3 narration track. This is stage 1 of the full pipeline (produce-video); use it standalone if you already have a silent video and just need the voice.

Requires ELEVENLABS_API_KEY (see .env.example at the project root — get a key at https://elevenlabs.io). No key, no narration — produce-video and film-demo both work fine without this skill (silent video only).

Quick start

pip install -r scripts/requirements.txt   # once

python scripts/generate_narration.py \
  --out demos/<slug>/narration.mp3 \
  --text "Here's how it works. Type your task, hit add, and watch it appear instantly."

Or --text-file script.txt for longer scripts. The script writes exactly one .mp3 at --out — no auto-naming, so it composes cleanly into a pipeline.

Defaults (override via flags or `demo.config.json` → `narration`)

Setting	Default	Flag
Voice	"Rachel" (search by name in your ElevenLabs voice library)	`--voice-name`
Voice ID (skips the name search)	—	`--voice-id`
Model	`eleven_v3`	`--model`
Stability	`1.0` (Robust — v3 scale: 0.0 Creative, 0.5 Natural, 1.0 Robust)	`--stability`
Output format	`mp3_44100_128`	`--format`

Pick a voice from your own ElevenLabs account (or a shared voice you've added to your library) and set it once in demo.config.json → narration.voiceName so every video in the project uses the same voice. --voice-name does a voices.search() by exact name — if you get "voice not found", the voice either isn't in your account yet (add it from the ElevenLabs Voice Library first) or the name doesn't match exactly; pass --voice-id instead to skip the search entirely.

Script-writing rules (these make sync easier later)

Plain sentences, one idea each. sync-narration aligns and cuts at sentence boundaries — a 40-word sentence covering three on-screen actions cannot be split apart later.
Write to match the pacing you'll film, not the other way around. ~150 words/minute is natural TTS speed — a 60-90s video needs roughly 150-220 words of narration.
No stage directions or brackets ("[pause]", "(click here)") — the model reads them aloud. Convey pacing with actual punctuation and sentence breaks.
Say what's on screen in the order it happens. sync-narration maps sentence order to beat order 1:1 — reordering later means re-aligning.

Gate before moving on

ffprobe -v error -show_entries format=duration -of csv=p=0 demos/<slug>/narration.mp3

Confirm the duration is plausible for the word count (~150 wpm). Regenerating costs API credits — only do it when the script text actually changes; reuse the existing .mp3 otherwise.

Deliverable

Report the output path, duration, voice/model used, and word count. If ELEVENLABS_API_KEY is missing, say so explicitly and stop — do not silently fall back to a different provider or skip narration without telling the user.