Faceless short-form: script, B-roll, voice, auto-post

Use this stack as a page: the faceless video pipeline.

A faceless channel is a factory, not a personality. The template is three stages: write a script, assemble a video without a face cam, publish to TikTok and Instagram on a timer.

“Viral ready” on a phone mockup is decoration. The real product is a repeatable 30–45 second object: script → pictures or generated clips → voice → captions → two platforms. If any of those is a one-off chat, you do not have a channel. You have a Tuesday night.

The stack we can actually name on this site:

1. Claude — script, trending-topic brief, on-screen text. ChatGPT if that is already home.
2. Picture / motion — Kling AI, Pika, Runway, Fliki, or stock you already license. Not a face.
3. ElevenLabs — the voice you reuse for months.
4. n8n or Zapier — auto-post and a thin analytics ping. TikTok/Instagram APIs are not matrix tools.

This is adjacent to the one-person creator stack, which is a look pipeline (Midjourney stills into Kling). Faceless shorts often skip a consistent character and live on B-roll, screens, and stock. Do not run both mental models on the same clip.

Why “just make a TikTok” in one tool breaks

Text-to-video models now emit native audio. That audio is still the wrong brand voice if you need the same narrator in 90 videos. Kling AI VIDEO 3.0 will talk. ElevenLabs is the voice you can version-control.

Fliki is the “script in, talking video out” shortcut: free row, paid from $10/month. It is honest about being a publisher’s assembler, not a film camera. Use it when the channel is listicles and stock. Do not use it when you promised cinematic product shots — that is Kling, Runway, or Sora (limited access, no public monthly price on our matrix).

Pika (free, paid from $10/month) is for quick motion tests. Runway (free, paid from $12/month) is for camera control. Pick one motion tool per video. Regenerating the same script in four generators is how faceless channels start looking like a model zoo.

Avatars (HeyGen, Synthesia) are a different product. If you put a digital human on screen, you are no longer faceless; you are in the HeyGen + n8n lane.

The three stages (and what each must not do)

Write owns the script, not the timeline

Claude paid from $20/month, long context. Output:

• 80–110 words, timestamps, hook in the first line.
• On-screen captions (short, not the whole paragraph).
• A shot list that does not require a face: screen recording, stock city, product macro, generated B-roll.

ChatGPT paid from $20/month is the swap. Perplexity can feed “what is trending in this niche this week” into Claude. Do not let Perplexity write the final voiceover; it is a research engine.

Do not ask Claude to upload to TikTok.

Assemble owns pictures and voice, not strategy

Motion: Kling from $10/month with a free credit pool, 15-second multi-shot on VIDEO 3.0. Stitch two generations if you need 40 seconds. Fliki if you are assembling stock + TTS in one canvas.

Voice: ElevenLabs, free row, paid from $5/month. One cloned your voice or a stock voice you will not abandon in week three.

Captions: burn them in the assembler or add them in the publisher. Auto-captions that disagree with the VO are worse than none.

Do not generate a new narrator voice every episode. That is how the channel loses the only consistency faceless video has.

Publish owns time, not taste

n8n (~$22/month cloud, or self-host free) posts when you say. Zapier ($19.99/month Professional after a small free tier) does the same with more consumer connectors.

Track performance in the native apps. A Zap that writes “views” into a sheet is fine. A model that “optimizes the algorithm” is fan fiction.

Respect platform posting rules, music licenses, and stock licenses. Unlimited auto-post of stolen clips is not an AI feature.

One episode, seven steps

1. Perplexity or your own notes: topic for this slot.
2. Claude writes script + shot list + caption lines.
3. Human cuts anything you would not say on a podcast.
4. Kling/Pika/Runway/Fliki produce the picture. ElevenLabs produces the VO.
5. Mux. Captions. Loudness. 9:16.
6. n8n posts TikTok and Instagram at the times you already know work.
7. Stop. Do not regenerate because the first hour of views was quiet.

What this stack is not

It is not a guarantee of reach. Faceless is a format. The algorithm does not owe you a factory.

It is not Suno as a voiceover. Suno is songs.

It is not Midjourney as the video tool. If you need a consistent illustrated look, go back to the creator stack and stay off auto-post until the look locks.

The page with the four matrix slots is the faceless video pipeline. ChatGPT can swap to Claude; Pika can swap to Kling AI.