Build a faceless channel that doesn't sound AI
Claude × Fish Audio
Hey, it's Adam. A faceless channel is a YouTube or Shorts account where the voice is the face — you never appear, but somebody has to hold the viewer's ear. My page runs on voiceover-led production every day, and the clone you heard in the reel is my actual voice, cloned on Fish Audio. This is the full workflow: Claude doesn't just write the script, it writes the performance — emotion tags dropped inline — and Fish Audio acts it out. Fish Audio partners with me on this content; the workflow is the one I'd use either way. You commented voice, so here it is.
Stop writing scripts. Write performances. 🎭
Faceless videos live or die in the first ten seconds, and the thing that kills them is almost never the footage. It's a flat voice reading a fine script. Viewers can't tell you why they swiped away — but their thumb already knew.
And the bar just moved: platforms are actively clearing out low-effort, mass-produced AI content from monetisation. A voice that actually performs — pauses, lifts, drops — is a big part of what separates your channel from that pile. The other part is saying something worth hearing. This guide handles the first; no tool handles the second.
The reframe that changes everything: a script tells the voice what to say. A performance tells it how. So you write the how directly into the text:
- 🧠 Claude is the showrunner — picks the niche, writes the script, and drops emotion tags inline: [confident] on the open, [calm] on the reveal.
- 🎙 Fish Audio is the actor — its S2.1 model reads those bracket tags and performs them. Paste the tagged script, get a delivery.
- 🧬 The voice is yours to choose — a library voice, a clone of your own, or one designed from a text description that has never existed.
Two tools. The whole channel's audio pipeline.
What you need ✅
- Claude — a paid plan (~US$20/mo). No code in this build; everything happens in chat and a browser.
- A Fish Audio account — the free tier gives you monthly credits and basic voice cloning for personal use. Two honest notes: a monetised channel is commercial use, and commercial rights sit on the paid plans — their current promo prices popular plans at 50% off through August 31, 2026 (their stated window — check the pricing page for what's live when you read this). Developers get the S2.1 Pro API free through the same date.
- A video editor you already use (CapCut is fine). Visuals are a separate guide — the voice is today's job.
- A shortlist of 2–3 topics you could talk about for fifty videos.
The showrunner prompt 📋
One prompt does the niche thinking and writes your first performance. Paste this into Claude:
You are the showrunner for a new faceless YouTube/Shorts channel. My interests and rough niche ideas: [your 2-3 topics]. STEP 1 — PICK THE NICHE Weigh each idea by: what advertisers pay in this topic (revenue per 1,000 monetised views decides the ceiling), whether I can make 50 videos in it without running dry, and how crowded it already is at the faceless level. Recommend ONE niche and a channel persona: a channel name, a one-line promise, and the voice's character (age, energy, attitude). Tell me what you rejected and why. STEP 2 — WRITE THE FIRST SCRIPT AS A PERFORMANCE A 60-second script, 130–160 words, written for the EAR: short sentences, one idea per sentence, nothing I could trip over reading aloud. Delivery rules: - Insert emotion tags in square brackets INLINE, exactly where the delivery should shift — e.g. [confident] on the open, [curious] into the setup, [calm] on the reveal, [warmly] on the close. - Tag the TURNS only: 4–6 tags total for the whole script. A tag on every line sounds unhinged. - The first 10 seconds must earn the next 50: open on the payoff or a concrete image. Never "in this video". - End with one line that makes the next episode feel inevitable. Then give me: a title under 60 characters the script fully pays off, and a 2-line description with the search phrase it targets. Never invent statistics. If the script needs a number, write [VERIFY: what to check] instead, so I source it before recording.
What a tagged script looks like coming back:
[confident] The most important machine of 1998 is in a landfill now. [curious] Not a computer — a beige box that quietly decided what your family watched, ate, and bought. [calm] It was called a TV meter, and only [VERIFY: how many] homes ever had one. [warmly] Stay to the end — because a box in a stranger's living room ran all of television.
That [VERIFY] marker is not decoration. Every number gets sourced before it gets recorded — invented stats are exactly the "AI slop" platforms are hunting, and they're also just wrong.
Test: read the script out loud twice — once flat, once obeying the tags. If the tags didn't change how you read it, they're decoration. Regenerate with a note about which turns felt wrong.
Let Fish Audio perform it 🎙
Try Fish Audio — free tier available now
Paste the whole tagged script into Fish Audio's generator — tags included, brackets and all. Model: S2.1 Pro. Pick any library voice for now and generate.
Listen for the turns, not the words: did the open land confident, did the reveal actually drop? Fixes, in order of how often they work: move the tag to the exact sentence where the shift should happen, split a long sentence into two, and only then swap the tag for a stronger one ([whispering] does more work than [quiet]).
Test: the ten-second test. Play just the first line — nothing else — and ask whether you'd keep watching. That single line is where the entire channel wins or loses. Spend disproportionate time on it.
Pick the voice: stock, cloned, or designed 🧬
Three doors, in order of commitment:
- 1A library voice. Fastest. The trade: other channels can use it too. Fine for testing a niche, not a moat.
- 2Clone yourself. Record a few minutes of clean audio — quiet room, no music, one continuous take, spoken the way you'd actually narrate. Upload, and the clone speaks your scripts with your texture and pacing. That's what you heard in the reel: my voice, without me recording a word of the script. Clone your own voice, or someone's who gave you written permission. Never anyone else's — that's not a grey area.
- 3Design one that never existed. Describe it — age, texture, pace, attitude — and Fish Audio builds a voice with no owner. This is the move for a channel persona you want to keep forever without it being tied to your throat. It's also the one that sounds, frankly, creepily human.
Whichever door: lock it. Save the voice ID, note the settings, and use the same voice for every episode. Viewers subscribe to a voice the way they subscribe to a face — changing it resets the relationship to zero.
Batch a week in one sitting 🔁
The channels that survive aren't the most talented, they're the most consistent — so make consistency cheap. Back to Claude, same chat: ask for the next five scripts in the persona's voice, same delivery rules. Generate all five audio tracks in one Fish Audio session while the settings are hot. You now have a week of episodes in an afternoon, and the visuals become a simple assembly job on top of a finished performance.
🔑 What keeps the channel alive
- 1Tag the turns, not the lines. 4–6 tags per 60 seconds.
- 2Write for the ear: short sentences, concrete images, no tongue-twisters.
- 3Every number gets a source or it gets cut. [VERIFY] is a workflow, not a suggestion.
- 4One channel, one voice, locked from episode one.
- 5Monetised channel = commercial use = paid plan. Don't build a business on a personal-use tier.
- 6If the platform has an AI-content disclosure toggle, use it. Disclosure isn't a penalty; getting caught without it is.
About that $10K a month ⚠️
- Some faceless channels clear five figures monthly. Most don't. What actually decides your number: the niche's ad rates (that's why the showrunner prompt weighs RPM), whether you ship on schedule for months, and whether the videos say anything. Treat $10K as the ceiling story that proves the model works — not the median outcome. Ranges, not promises.
- The platform crackdown on mass-produced AI content is real, and a performed voice doesn't exempt you — it just stops you failing the first filter. Substance passes the second one.
- The dates matter: the free S2.1 Pro API window and the 50%-off plan promo both run through August 31, 2026 — if you're reading this after, check current pricing before you plan around it.
- Clones inherit everything — your accent, your pacing habits, the way you rush endings. That's the charm and the limit. If your narration habits are bad, the clone is faithful to them. Fix the read, then clone.