My case: launching YouTube channels solo — part 1, from a music label to seven channels and my own analyser

This is the second live log in the series. The first one was building a one-person AI music label. This one is what the label turned into once the music existed: a set of YouTube channels run by one person, with my own analyser, my own content plan and my own list of open questions. Same rules as everything on this site — structure first, no rituals, no numbers I can’t show.

Part 1 — how the label became channels

Streaming pays fractions of a cent. That was clear from the AI-music economics, so the label’s plan always treated streaming as presence and YouTube as the audience layer. The first channels were therefore the obvious ones: one channel per artist persona, publishing the tracks as clips — a single still scene per song, film grain and a flickering streetlamp for life, and the lyrics running as word-by-word karaoke over the top. Cheap to make, and the karaoke text is not decoration: it is the reason someone stays for three minutes instead of swiping.

Then the same production line started spawning other things, each as its own channel — and each channel is its own combination of category, language and visual style. Not one house style: a drawn caricature here, film-grain noir there, neon graffiti, photoreal landscapes, warm stills with text. Not one language: some channels are in Russian, some in English. Those differences are deliberate — they are the variables the whole set is testing, so the same production line can show which combination the feed picks up:

• Education in 30 seconds — short musical facts: a rap-pop track about a historical figure, a drawn full-length caricature of the person, dates as stickers, the lyrics as karaoke. “All of world culture in short musical stories” was the concept; the format is a Short; the language is Russian.
• Silent stories — wordless character shorts with a fixed cast, told entirely in composition and on-screen captions of one or two words. Two variants, two styles: monochrome noir with a couple, and neon graffiti with a crew of five. English captions.
• Heartfelt stories — a voice reading a short, warm story over generated stills; captions land under the words as they’re spoken. No faces in frame — an artistic choice that is also a technical one, because image models don’t hear “no face”, they only obey composition.
• Nature and landscapes — walking-into-the-frame footage where the person is always far away or seen from behind.
• News digests — short reaction formats built on the events already being tracked for my news project.
• Games — gameplay of my own prototypes from the game-studio case, released on the channel before they reach any store, which turns “unfinished” into “see it first”.

Seven channels in total by mid-September. The label idea — breadth as inventory, then grow what moves — applied to channels instead of tracks.

How the niches were actually chosen

Not by market analysis first. The filter came in this order:

1. Can I produce it at near-zero marginal cost with what I already have? A story Short costs me 20–45 cents in image generation; a nature episode under a dollar; a news digest nothing. Anything that needed a camera, a face, an actor or a studio was out before it was ranked.
2. Does it fit what I care about and can sustain? A channel is a promise to publish for months. My own interests and a written list of excluded topics sit in a config file that the idea generator reads before it ranks anything — an idea I can’t or won’t make is not an idea.
3. Does the data say the niche has a ceiling? This came last and mostly confirmed or killed things I already suspected (more on the language finding below).

The identities were done with an AI chat assistant and an image model: avatar, banner, cover template and a one-paragraph visual doctrine per channel — each one different, each one fixed once written (“caricature with a big head, period costume, no children’s-cartoon look, almost static with a little motion” for one; “one static courtyard, only the streetlamp flickers, digitised-film grain” for another) — saved into that channel’s brand folder that every future episode is checked against. Style is consistent within a channel and deliberately different between channels. Identity was the cheapest part of the whole thing and the one that never needed redoing.

Dozens of coaches → one guide

Before publishing anything I did the thing every beginner does: watched a lot of YouTube growth coaches. Then I did the thing most don’t — collected every tutorial, transcript and screenshot I’d found and had an AI assistant distil them into a single standard, deduplicated and stripped of the motivational filler. It became three working documents that the whole line runs on:

• A hook standard. The first three seconds have to do three jobs at once: stop the scroll, make a promise, open a loop that only the ending closes. The phrase itself: involvement + urgency + value + intrigue, eight words or fewer, spoken aloud and mirrored on screen one or two words at a time. Six hook types, rotated so no two consecutive episodes use the same one. A banned list: “Welcome”, “Today I’m going to”, “Did you know”, any clickbait adjective.
• How YouTube evaluates a new channel — five stages: it first decides who you are from the first uploads; then hunts for your audience with small test groups; then compares your videos against each other; then maybe gives one or two a wide push; then watches for regularity. Most beginner “stuck” problems are stage 1–2: mixed topics and generic tags mean the channel never manages to introduce itself.
• A channel canon for both Shorts and long-form: finish the channel page before the first upload (avatar, banner, description, trailer, playlists, verification level); don’t touch title or thumbnail for 48 hours; read the retention graph, not the click-through rate; the long-form skeleton — 15-second hook, chapter map, 60–120-second blocks with open loops, a pattern interrupt every 4–5 minutes, answer the hook at 85–90 %, 20 seconds under the end screen.

None of that is original. What’s mine is that it is one document, not forty videos, and every episode is checked against it before upload.

The analyser — because coaches don’t show their data

The coaches’ advice all sounded plausible and none of it came with numbers. So I built the thing I wanted to exist: a script that collects Shorts from public YouTube pages and the free Data API, and for each one computes lift — log10(views) − expected(subscribers, channel age, niche). Lift over the channel’s own norm, not raw views, because a million views on a million-subscriber channel is a Tuesday.

The corpus reached 6,566 Shorts from 646 channels. What survived on that sample (an earlier pass on 500 videos produced conclusions that later reversed — small samples lie, and I logged that too):

• Channel size matters less than you’d think. A channel ten times larger gets about 3.6× the views on a Short, not 10×. A hit without an audience is the feed’s normal mode, not a miracle.
• Age stopped predicting anything. A Short gets its views immediately and barely accumulates after.
• Length is the strongest controllable lever. Under 8 seconds: a clear positive. 61–90 seconds: a penalty. 91–185 seconds: a heavy penalty.
• Niche is a ceiling you pick once. Hobby and style, film and animation, people and blogs sit above zero. Gaming — my most natural niche — is measurably below. Travel and news are far below.
• Language is a hard ceiling — and a variable. Estonian, my home market, carries the heaviest penalty in the whole model, so there is no Estonian channel. Spanish and Portuguese sit above zero; English and Russian both near it — the data doesn’t separate them, so the channels do: the music and education channels run in Russian, the story, nature and news channels in English, and the tracker compares them.
• A link in the description correlates with a penalty. Challenge- and story-shaped titles with a small bonus.

And the finding that changed how I spend my time:

So the next version of the analyser went inside the frame: download the video at low resolution, run it through ffmpeg into raw frames, and measure structure. Paired sampling — one author’s best video against the same author’s worst, so only structure differs. What held in the gaming niche over 58 pairs: the winners were shorter than the same author’s losers, had fewer cuts in the first third (one against four — the opposite of “cut fast”), no more than one cut in the first three seconds, and the first cut slightly earlier. That template now sits in every episode brief, rescaled to the target length.

A method lesson that cost a day: two collection paths (a downloader and the API) return different values for the same fields, and a mixed corpus produced spectacular fake effects — the model was measuring how the data was collected, not the videos. Fix: tag every record with its method and fit one method at a time.

A second script does the reverse job: paste a channel URL and get a report — its norm, its outliers, Shorts and long-form separated, its rhythm and length distribution. I ran it over popular channels in the formats I want to learn from, especially channels that grow on 20-minute-plus videos, because that is where the watch-hours are.

The model — two layers and a loop

The analyser is the technical base: markers that work for everyone. On top of it sits my layer: category, idea, story, filtered by my interests and my production limits. The output is a semi-finished brief — either ready to shoot or one more pass away. And the loop closes with a tracking script: every upload logged, views measured at 24 hours, 72 hours and 7 days, and a report on what to scale and what not to repeat. Without the loop the tool is open-ended and therefore useless.

The content plan that came out of it, per channel:

• A living spec file — format, cast rules, visual doctrine, hook rotation — updated after every correction I make, so the same mistake can’t repeat.
• A release schedule across all channels: one episode per channel per day when the budget allows, with two or three finished episodes in reserve per channel, because YouTube’s fifth stage is regularity and a burst followed by silence throws you back to stage two.
• Cost per episode written next to each one, so the pause is deliberate when generation credit runs out — which it has, twice.

What the first weeks actually showed — by direction

Every direction is its own experiment, so it gets its own data. Numbers are what my channel dashboards show as of mid-September; where a direction is too young to have a number, it says so instead of getting one.

Music clips (artist channels, Russian). The one thing the feed has kept distributing is a full-length clip, not a Short. One three-minute clip of a Russian-language persona was in recommendations from the moment it went up: about a hundred plays in the first twelve hours, and past a thousand since — almost all from suggested videos and search, none from anything I did. The same persona’s other tracks sit far below it, on the same channel, in the same style, with the same voice. That is the analyser’s finding playing out live: what differed was inside the track and the frame, not the metadata. It is also the direction closest to YouTube’s stage 3 — the feed is comparing that channel’s videos against each other.

Education Shorts (Russian). The channel introduced itself well, then I broke it. The first nine Shorts were each picked up by the feed within a couple of hours. Then I uploaded eight files in one batch and scheduled them two a day — and the next two published got zero feed distribution in five hours, only my own devices. My reading: a channel that suddenly looks like a same-shaped conveyor gets throttled. Regularity means a rhythm, not a dump. Two other fixes came from here: I had been tagging historical figures as “People & Blogs” and science topics as “Education” on the same channel — now one category per channel, sub-series told apart by title, playlist and tags; and after comparing sub-series, the channel narrowed to historical figures only. Twenty-one episodes made, the channel is on YouTube’s stage 2.

Shorts across all channels. The best Shorts reached roughly 1,200 views; the typical Short sits in the tens and low hundreds. Real numbers, weeks old, no audience underneath — exactly the regime the analyser was built for.

Silent stories (English, two channels). Two episodes published on each, four more finished and in reserve on the noir one. Too young for a verdict: views in the tens, both channels on stage 1 — YouTube is still deciding what they are. The open question is voice: a synthetic narrator is the next test.

Heartfelt stories (English). The most-published direction — nine episodes out in two weeks at 24–45 cents each — and the one where the production rules got hardest: no faces, hero checklist per frame, a warm one-line caption over the whole episode. Views low hundreds at best; stage 1–2. The 7-day measurements on the current run decide whether it earns more episodes.

Nature (English). Five episodes out, under a dollar each; the direction came from scrolling my own feed, not from the analyser — and the analyser’s travel-niche penalty is the reason it stays small until it proves itself. Stage 1–2, no number worth quoting.

News digests (English). One test episode published on the new format; the earlier digest Shorts from June–July predate the analyser and the hook standard, so they are the baseline, not the test. The analyser’s news-niche penalty is heavy; this direction survives only because the events are already being tracked for the news project, so the marginal cost is zero.

Games (English). The channel exists, the UGC-style intro pipeline works (a generated desk-and-phone scene, a few seconds of motion, the real gameplay tracked onto the screen), the first five intros are built — no episode is published yet. Gaming is measurably the hardest niche in the corpus, so it goes last.

All of it together. Every channel sits at YouTube’s stage 1–2 as of mid-September. One clip past a thousand, Shorts up to about 1,200, everything else in the tens and low hundreds. No channel is near Partner Programme thresholds, and the thresholds are being raised again next year — which only reinforces the label’s original thesis: YouTube is the audience layer; the asset is the catalog and the list. Revenue: zero. Expected. Logged.

The problems that turned into rules along the way: the music tool’s download limit dropped to twenty a month under new terms, which turned lyrics into the thing to write in advance and downloads into the scarce resource; generation credit runs out and pauses the line — so the schedule shows cost per episode and free work (analysis, renaming, planning) fills the gap; my own test views polluted early numbers, so the tracker ignores the first day; the shared hosting I use for websites can’t run the analysis (no Python, no ffmpeg), so everything computes locally and the web dashboard is only a window onto it; and YouTube’s synthetic-content disclosure needs reading properly rather than ticked by reflex — it covers specific cases, not “made with AI”.

What I’m testing next

• A measuring channel, not a growth channel. One channel on a neutral account, tied to no brand, with zero external traffic, one publishing hour, nothing ever deleted. Its content is a procedural, silent animation rendered in the browser — free, infinitely variable, every parameter a number in a config. Phase 0 is fourteen videos built to the same spec, to learn how wildly results vary with nothing changed at all. Only after that: one variable per phase, eight videos a group — length (8 / 15 / 25 / 40 / 60 s), seamless loop versus not, on-screen text none / caption / word-level. Inside your own channel you see what no outside dataset has: retention, swipe-away rate, repeat views, impressions.
• Voice on the silent channels. Picking a synthetic narrator for the wordless story formats and running one episode each way, because 60 % of viewers watch without sound and the other 40 % may be leaving.
• A long-form channel built from the reports. The channel analyser has been pointed at channels that grow on 20-minute-plus videos; the next channel is designed from those reports and the canon’s long-form skeleton, not from a Shorts habit.
• My own voice, dubbed. For this site’s channel the plan is one recording session in my native language, dubbed into English in my own voice, over screen recordings — one session a month covering a month of videos.
• Chaptered descriptions as a variable. Same channel, alternating episodes: plot summary with chapters versus a one-liner, everything else fixed — to find out whether the clip’s early push was the description or the track.
• Frame-level structure in every brief, then checking whether episodes that follow the template outperform the same channel’s episodes that don’t. That’s the only test that matters for the analyser itself.

Why this maps to everything else on this site

Strip YouTube out and it is the same solo playbook as the label and the games: breadth as inventory, a production line that makes the marginal unit nearly free, one owned asset underneath (here: the catalog and, eventually, the list), a measuring loop instead of opinions, and no numbers until they are real. It is also the first case where building the instrument turned out to be worth more than any single output — the analyser told me in a week what forty coaching videos couldn’t: pick niche and language once, then work inside the frame.

The log — results as they land

• 2026-09 (part 1, this post): seven channels publishing across music, education, silent stories, heartfelt stories, nature, news and games; identities, specs and release schedule in place; hook standard, five-stage model and channel canon written; Shorts analyser fitted on 6,566 videos / 646 channels with a frame-level structural template; tracking loop live. Languages split Russian / English, one visual style per channel, none shared. Best clip past 1,000 plays from recommendations and search; best Shorts ≈1,200 views; results logged per direction above; all channels at YouTube stage 1–2. Revenue: none. Next checkpoints: 7-day measurements on the current run, the measuring channel’s phase 0, first long-form channel design.

Updates get appended here with dates as real numbers arrive — wins and failures both.

The general playbook this case tests: how to build a faceless video channel solo and what a YouTube channel actually pays a solo. Where the direction sits on the ladder: the boards.

Part of the personal log. The thinking this experiment belongs to: living on the border of the synthesised world.

The production line in this log, as recipes you can run: brand identity · creatives and thumbnails · a video without a camera · publishing on a channel.