Polli Video

Single command, no adelay math, perfect sync. Requires BlackHole 2ch and default system output set to it (Option+click menu-bar speaker, or SwitchAudioSource -s "BlackHole 2ch").

See example/render.sh — archives old renders, captures while VHS runs, muxes trimmed to video length with a 5s audio pre-roll shift (ffmpeg startup + API roundtrip offset) and a 3s audio fade-out at the end. Copy and adapt.

Capture tool: sox over ffmpeg avfoundation. ffmpeg dropped samples on long recordings (choppy audio) even with -thread_queue_size 4096, -async 1, nice -20. sox + coreaudio is stable end-to-end, addresses device by name ("BlackHole 2ch") instead of a flaky index, and needs no post-trim:

AUDIODRIVER=coreaudio sox -q -c 2 -r 48000 -t coreaudio "BlackHole 2ch" \
    -c 2 -r 48000 -b 16 captured.wav 2>/tmp/sox-capture.log &

Music gen must use --play so BlackHole captures it. Switch default output back to Speakers after.

Pipeline B — post-hoc ffmpeg overlay (fallback)

When BlackHole unavailable. Pre-generate narration + music, overlay with manual timestamps:

vhs demo.tape   # produces demo-silent.mp4
ffmpeg -y -i demo-silent.mp4 -i speech.mp3 -i music.mp3 \
  -filter_complex "[1:a]adelay=16000|16000[narr]; \
                   [2:a]volume=0.18,atrim=end=75[bg]; \
                   [narr][bg]amix=inputs=2:duration=longest[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac demo.mp4

adelay=16000|16000 = ms, left|right (must be stereo pair). Tune to tape timestamp where narration should hit. Ducking: 0.12–0.20 under narration, 0.25–0.35 music-only. Never pass -shortest — truncates video to shortest audio.

VHS tape conventions

• Set Shell "bash" is mandatory. zsh + long Type lines triggers a VHS command-concatenation bug: two consecutive long commands get glued onto a single prompt line and the Enter between them is lost. Root cause is zsh's ZLE + VHS's wall-clock-timed CDP keystroke injection (no prompt-wait primitive). Bash's simpler line editor has no such race. Not fixable with Wait+Line (regex fails on colored prompts), precmd hooks (VHS #691), longer Sleeps, or Ctrl+L substitution. Just use bash.
• 960×720 or 960×640, Menlo 16pt — reads on mobile, fits polli gen text --model X "...".
• Absolute path in Hide-block cd; relative path in Output (parser splits on /).
• Always export FORCE_COLOR=1 in Hide — VHS pty fails isTTY, chalk strips colors otherwise.
• Scene titles via ANSI from shell (no VHS primitive). \033[1;38;5;141m = polli purple.
• announce() pattern: one function that prints the bold title AND fires backgrounded TTS. Every later scene is a one-liner. See example/demo.tape.

Canonical bash Hide block (copy verbatim — fixes job notices, heredoc prompts, prompt trailing-line):

Set Shell "bash"
Hide
Type "cd /absolute/path/to/working/dir"
Enter
Type "export FORCE_COLOR=1"
Enter
Type "export PS1='» '"
Enter
Type "export PS2='» '"
Enter
Type "PROMPT_COMMAND='echo'"
Enter
Type "set +m"
Enter
Type "clear"
Enter
Show

• PS2='» ' — hides heredoc/continuation > prompts that leak on wrapped Type lines.
• PROMPT_COMMAND='echo' — trailing blank line between scenes so output doesn't crash into the next prompt.
• set +m — disables job-control, suppresses [1] 12345 notices from & backgrounding.
• Wrap backgrounded commands in a subshell anyway: ( cmd & ) never prints a PID even in interactive shells.

The announce function

One-liner per scene: prints a bold purple title AND fires backgrounded TTS. The sed strip removes ElevenLabs emotion cues ([whispers], [excited], …) from the printed title while keeping them in the spoken audio.

announce() {
  printf '\n\033[1;38;5;141m»» %s\033[0m\n\n' "$(sed -E 's/\[[^]]*\] *//g' <<< "$1")"
  ( polli gen audio --play --output /tmp/narr-$$-$RANDOM.mp3 "$1" >/dev/null 2>&1 & )
}

Usage:

Type `announce "[excited] first — we need some backing music"`
Enter

ElevenLabs emotion cues (must use ElevenLabs-backed voice, e.g. default sage): [whispers], [excited], [confident], [curious], [sighs], [laughs]. Place inline — the TTS engine interprets them; the sed strip keeps the on-screen text clean.

Why subshell ( cmd & ), not bare & or disown:

• Bare & in a non-interactive bash (VHS) still prints [1] 12345 — even with set +m, occasional leakage.
• disown after & still shows the job-start notice.
• ( cmd & ) spawns the background job inside a subshell that exits immediately, so no job record is ever attached to the parent shell — nothing to print.

Gotchas

Voice selection

Default: sage. Others: fin, callum, onyx, rachel. Preview:

polli gen audio --voice <name> --output sample.mp3 "test line" && afplay sample.mp3

Full list: polli models --type audio --json | jq '.[].voices'.

Music (elevenmusic)

Deterministic cache — same prompt+duration = same bytes. Always --instrumental. Match duration to ffprobe demo-silent.mp4. Example prompt: "lofi instrumental hip hop, warm analog, jazzy bass".

Demo design principles

• Open silent with polli --help — orients without narration.
• One audio moment per scene, not many.
• Hold scenes: help ~4s, streaming text 10–20s, ASCII outro 15–17s.
• Payoff produces something reusable (the post, an image, a doc).
• Streaming > buffered — lean into polli's default streaming reveal (drop --no-stream when used in a recorded tape so the audience sees generation happening live).
• End with a punchy ASCII wordmark outro (see "Streaming outros" section).
• 960×720 or 960×640 beats 1920×1080 on mobile feeds.

Brand voice for generated content

See social/prompts/tone/linkedin.md, social/prompts/brand/about.md.

• Dry, information-dense, anti-corporate.
• No "excited to announce", "game-changing", hashtag spam.
• Plain text for LinkedIn (markdown renders raw).

Generating the announcement post (lessons)

Model selection (ranked for this task):

• claude-fast — pick this for live demos. Consistent ~4s latency, never leaks scaffolding in this prompt shape. Voice slightly flatter than glm but reliability wins when a Sleep 30000ms hold has to cover the call.
• glm — best voice quality when it works. But latency is wildly variable (12s first call → 31s next → sometimes >60s); not safe behind a fixed Sleep budget. Use for offline one-shots, not live-tape renders.
• kimi — strong rhythm, hacker-genz feel. Solid backup.
• openai — leaks prompt scaffolding verbatim (e.g. sign-off:, 110w, your own line-1 hook), overshoots word count. Avoid.
• gemini-search / perplexity-fast — web-search adds citation markers [1][2]; hallucinate features, inflate corpo voice. Avoid for this.

Call flags:

• Use --no-stream. Streaming + > file hangs on long prompts; --no-stream returns in 3–5s.
• Never wrap polli in timeout N — it breaks the stdout flush and always kills the call.

Input context (what to pipe in):

• polli --help only — leanest, forces the model to describe the command surface without repackaging README bullet points.
• Feeding the README causes models to repeat README-specific phrases ("humans, AI agents, and everything in between") and hallucinate flags (--budget etc).
• --help limitation: doesn't expose npm package name → model guesses polli-cli. Patch install line manually.

Prompt shape (compressed, declarative, no imperative wordiness):

• State the task (announce polli on LinkedIn), not a procedure.
• Voice in 1 line with · separators. Let the model interpret rather than listing banned phrases.
• Don't put counted claims in the prompt (7-verb CLI) — every model echoes them. Just list the verbs.
• Don't say "scroll-stopper" or "your own line-1 hook" — models paste those labels literally.
• Keep no md · no 🚀 · no corpo as the only negative constraints. More bans → defensive, boring output.
• Ask for sign-off not CTA.

Final prompt used (LinkedIn launch post):

↑ polli --help for polli CLI.