Skip to content

Claude Code skill that makes launch videos, ads and music videos from one line: idea, script, visuals, voice, music, edit and review, then a review room where your notes make the next take

7.0 KB
Raw Blame

The voice: casting, recording, cutting

A narrator is a choice, not a default. Use one when the film has to explain something a viewer cannot read from the pictures, and write the words first (references/words.md). The goal is a voice that sounds like a person talking, not a voice that sounds generated. Voice generation costs credits on most plans: say the price first.

Which account

Not at the start and not in the pitch. Only at a moment of need: a voice is a must for what they asked (a narrated film, a voice-over ad), the first voice is about to be made, narration comes up in the middle of a film, the person is stuck without a voice, or they ask to take the film higher with one. Then check whether the person has ElevenLabs: a connector in the session, a logged-in tab in Claude in Chrome, or an API key they have set. If they have it, offer to use their own account and say what one narration will cost from their plan. If they do not, offer the free way to start: they can make a free account at https://try.elevenlabs.io/take-one, with this line in the same message: "This is Tal's partner link (he made this skill). If you sign up through it he may get a commission, at no cost to you; elevenlabs.io works the same without it." They sign up and log in themselves. Offer the link once; after a no, go on with the other ways (they record the lines themselves, or no voice, services.md). A film for a business may need a paid plan for commercial use: check ElevenLabs' current terms and say so.

Casting

The voice director's job is to bring the right three voices, not any three. The last choice is made by ear, and only the person has ears.

The voice brief, first. Take it from the strategist's brief and write it in a few lines before opening any voice library:

  • who is listening, and where (founders on a phone, parents in a feed, buyers in a meeting);
  • what they should feel when it ends (trust, appetite, calm, a smile);
  • the brand's tone in three words, from its own site and films;
  • so: an age range, an energy, a pace, how warm, which accent, and what it must not sound like (an announcer, a hype read, a machine).

The shortlist. Search the service's library by its own labels and descriptions (age, accent, use: conversational, narration, advertising) and pick three that fit this brief and differ from each other in one clear way (energy, age, warmth or pace). Three versions of the same voice is not a choice.

A new search for every film. The shortlist comes from this film's brief, never from a list of favourites and never from the last project. There are no default voices in this skill. Two films for different audiences should not end up with the same three candidates; if they do, the brief was not used. The one exception is a brand that already has its voice (3b): offer that voice as one of the three, next to two new ones, and say so.

The samples. Record the same two real lines from the script with each voice. Then check what can be checked without hearing: every word is there (transcribe it), the pace in words a minute fits the picture, no line starts clipped. Say which one you would cast and why, in one line, marked as yours.

  1. The person picks the voice by ear. The maker cannot hear. Send the two or three short samples of the real lines, not a demo sentence, and let them choose.
  2. Never use the person's own cloned voice unless they ask for it.
  3. A second voice (a character, an ad inside the film) is a second casting with its own samples. 3b. Keep the choice: write the chosen voice, its model and settings into the project, and into the brand's saved profile if there is one, so the next film for this brand starts from it.

Recording

  1. Use the service's newest speech model; it is usually the one that sounds like talking.
  2. One take for the whole narration, so it is one person in one mood. A different voice gets its own take.
  3. Few direction tags: one per change of mood, in plain words. A tag on every line sounds acted. If the service adds tags by itself, read them back before generating.
  4. Put a pause between lines, so the take can be cut at real silences.
  5. A line added later is recorded together with the line before it, and only the new line is placed. The approved take is not touched.

Checking and cutting

  1. Transcribe the whole take: every line there, in order, no tag spoken aloud.
  2. Check its length against the picture before building on it. A more acted read can run a quarter longer.
  3. Cut silence to silence, never by word times alone, and transcribe every cut line on its own before it goes in. scripts/vo.py does both and stops the build on a line that does not match its text:
python3 <skill>/scripts/vo.py vo.json   # takes + lines + start times in; audio/vo.wav and assets/vo.js (when each line starts) out
  1. Words on screen follow the placed voice (read assets/vo.js), never the other way round.

In the mix

  1. Voice sits 10 to 17 dB over the music while it speaks. A voiced piece inside the film needs a deeper dip.
  2. A sound effect on the same instant as a short word hides the word: sound first, word a beat later.
  3. Check the true peak after mixing (scripts/audio-check.mjs).
  4. Space the lines evenly: about 0.3 s between one sentence and the next in a list, never 0.1 in one stretch and 0.5 in another. Lay the voice out first, on its own, then put each picture's cut at or just before its line (up to 0.4 s before is fine; the picture leads, the voice names it).
  5. Let a drop land before the next line: give it half a second with no voice over it, or it is heard only as a dip.
  6. A second voice inside the film (an ad's own narrator) gets a lighter dip under it than the narrator, about 8 to 11 dB, so the music under that piece keeps its weight; check each of its lines still clears the music by 7 dB or more.
  7. Transcribe the final mix (whisper-cli on the mixed file) and read it: every line must come back as written. It is the one check of the mix that needs no ears.
  8. The check is not optional when you build your own pipeline. On an 8-minute documentary the maker re-spaced a long take with its own script instead of scripts/vo.py, cut at whisper's line times, and skipped steps 11 and 19. Fifteen cuts landed inside words ("Septem-ber" chopped at 0:08). Three fresh reviewers and its own taste review missed it, because none of them can hear and the sound report said nothing was wrong; the owner heard it in the first ten seconds. Whatever tool cuts the voice: choose each cut as the pause where the audio before it ends with the previous line's last words and the audio after it starts with the next line's first word, then transcribe every placed line in the final voice track and compare first and last words before any render. Put that line-by-line result in the reviewer's pack, so a reviewer who cannot hear can still see a failed cut.

Working in the person's account

  1. Their page is theirs. Note the text, voice and model before touching it, and put all three back at the end.
  2. Say what was spent, even when it came from a free pool.