I run a faceless YouTube channel. For most of last year the production looked the way it looks for everyone: a folder of half-finished scripts, a browser full of "how to write hooks" tabs, and an editing session that started at 9pm and ended in regret.
The usual fixes didn't fix it. A course taught me theory and then left me alone with an empty timeline. The auto-video tools produced clips in thirty seconds that looked exactly like everyone else's thirty-second clips, because they came from the same templates.
So I tried a third thing. I wrote my entire production process down as plain text files — every rule, every checklist, every "never do this again" — and handed the folder to Claude Code.
Claude read it and became the studio.
The folder
Here is the actual structure, 39 plain-text files:
ChannelForge/
├── START.md · the installer Claude executes
├── setup/ · environment check, install rules, an interview
├── pipeline/ · the map + 6 phase playbooks
├── skills/ · 8 skills: niche scout, topic research,
│ titles & hooks, thumbnails, script writer,
│ image prompts, publish packages, shorts
├── scripts/ · 9 tools: word-level audio sync, asset
│ control, timeline assembly (2 paths)
├── templates/ · channel profile, checklists, tracker
└── docs/ · FAQ, troubleshooting, MAKE-IT-YOURS
No API keys. No servers. No automation graphs that break at 2am. Files a language model can read, plus the subscriptions I already had: Claude, an image generator, a voice service.
The install is one sentence: "read the ChannelForge folder and set it up for me." Claude checks the machine, asks eleven questions about the channel — niche, style, video length, which image provider I pay for — and writes the answers into a profile that every later step reads.
How a video moves through it
Phase 1 — the niche. The niche-scout skill has a rule I stole from my own failures: no numbers from memory. Claude verifies every niche candidate with live web search — real competitor channels, real view counts, and a monopolization check: do small channels actually get views here, or do three giants own every search result? Each candidate becomes a card with a verdict, GO or NO-GO, and sources linked. I pick from verified options.
Phase 2 — the concept. Topic, title variants, and the thumbnail brief. The thumbnail comes before the script, which sounds backwards until you watch a channel die at a 2% click-through rate. The thumbnail is a promise; the script exists to keep it.
Phase 3 — the script. Retention rules are written into the playbook: the hook lives in the first 30 words, every block ends on an open loop, no block is allowed to be a pure fact list. A draft passes a 7-point checklist before I ever see it. When a check fails — say, a middle block that reads like a fact list — that block gets rewritten before the draft reaches me.
Phase 4 — assets. Claude writes image prompts for my provider, in my channel's visual canon, with a consistent character across scenes. I generate the images in my own account — semi-manually, by design. No API keys to leak, no middleman charging me per image.
Phase 5 — the edit. A script transcribes my finished voiceover word by word and maps every image to the exact frame where its phrase begins. Then it either assembles the timeline inside DaVinci Resolve Studio directly, or generates a timeline file that opens in the free DaVinci, Final Cut, or Premiere — every image on its frame, voice and music on their tracks. It reads the result back and reports drift — "no desync," or a list of scenes to nudge.
Phase 6 — publish. Title, description, tags, chapters — built by rules, not vibes. The system re-checks current YouTube policy by live search at publish time, because platform rules change faster than any file I could have written.
The gates
The part I trust most is the part that says no.
Before anything ships, I watch the first 60 seconds of my own video the way a stranger would, and answer one question: would I swipe? If the answer is yes, the hook gets rebuilt — text, images, or motion — and the video does not go out that day. A rebuilt hook costs an evening. A burned first impression costs the video.
There's also a stop-valve at the channel level: two flops in a row and the system forces an analytics review before it lets me produce the next video. That rule exists because I once spent a month producing into the void, and the pipeline without a brake happily let me.
What it doesn't do
Honesty section, because this niche is drowning in promises.
It needs your own Claude Code subscription. Images and voice are generated in your own services — the pipeline prepares exact prompts and filenames, you press the buttons. And it makes no promises about views or income, because nobody can promise that: the audience decides. What the system guarantees is a disciplined process — checked facts, quality gates, an honest analytics loop.
The economics are simple to check, though. Freelancers charge $110–340 per faceless video. The SaaS automation stacks bill $47–180 every month, forever. A video through this pipeline costs me $1–5 in generations on subscriptions I already had.
Owning the studio
The thing I underestimated: because the whole system is readable text, it rebuilds itself on request. "Make videos 20 minutes long." "Add a TikTok export step." Claude rewrites its own playbooks. There is no roadmap and no feature request queue — there's whatever you say next.
I packaged the folder — the pipeline, the 8 skills, the 9 scripts, the checklists with scars behind them — into a product called ChannelForge. It's $19, once, and it runs as many channels as you can feed. Everything I described in this article is in files you can open and read fifteen minutes after you download it.
If you've been stuck between "another course" and "another subscription," the folder is here: https://channelforgehq.github.io/
Top comments (0)