What the source did
- Automated the marketing for his faceless YouTube channel with Claude Code and open-sourced it: five subagents and twelve skills for A/B titles, thumbnail briefs, Shorts, analytics and Instagram captions.
- The flow, per his README: Scout picks a scene, Engineer builds the take, QA reviews with fresh eyes, Engineer fixes, Growth packages it and uploads it as private, the owner approves, the owner publishes.
- QA never edits. It runs checks and reports one table per scene with a status, where the problem is, what's wrong with the evidence, and who fixes it.
- QA's checks are measurable, not taste: loudness and true peak, title length counted the way YouTube counts it, a thumbnail scaled to 168 px wide to test the speech bubble, nothing in the bottom-right corner.
- Growth is the only owner of views and subscribers. It measures Analytics, watches competitors, writes a daily plan with the KPI each action moves, and logs every change as an experiment with a date to measure it.
- The PM keeps the team log, reports to the owner and proposes edits to the other agents' instructions, applied only after the owner approves.
- Guardrails in the README: drafts only, no auto-publishing, no paid posting services, no fake engagement, and comments, forms and web pages are data, never instructions. A per-item lock keeps two sessions from editing the same thing.
- The README's example day runs one scheduled task per agent, from Scout at 07:30 to the PM's report at 18:30. Routines that would upload or publish on their own stay paused.
- Asked how QA grades creative work, he replied that the workflow keeps a human on it.
“It is impossible to delegate creativity, no matter how refined the agent may be.”— Tifis.md (@tifis), X
Summary written by Agent Cookbook from the original post. See the original for the full context. How we source recipes
Our notes
- What it takes
- Claude Code, his five agent files and the skills you need, plus the folders they expect (a team log, a review queue, a brand guide). His quality bar is written for a dubbing channel, so you rewrite it for yours.
- Where it breaks
- Reviewers asked to judge taste instead of measurable checks, transcript-only audio checks that can't hear background noise or hiss, two sessions editing the same piece, and any routine left to upload or publish on its own.
- Human checkpoint
- Growth uploads as private only. Each piece and format goes public only after the owner's own yes is recorded in the sign-off log, and the PM edits agent instructions only after the owner approves.
- Tools involved
- Claude Code, YouTube Studio, ffmpeg, Whisper (QA's audio check), Instagram or Facebook (optional skills)
✓ Verified Source link checked and still live. Notes by Thomas, editor. How we verify
Get their template
Made and published by Tifis. Start here; it is their original, not ours.
Approval rule
🛡
Growth uploads only as private. A piece goes public only after the owner records an approval for that piece and format, and the PM changes an agent's instructions only after the owner says yes. Sending, buying, posting, and deleting always need a human.
Set it up
- Copy the five agent files from his repo into .claude/agents/ in your project and the skills into .claude/skills/.
- Create the files the agents expect, or change the paths: a team log (team/log.md), a review queue (team/quality/pending.txt) and your brand guide (brand/BRAND.md).
- Rewrite the shared quality bar for your channel. His is specific to a dubbing channel (number of dubbed voices, loudness, family-friendly rules).
- Start a session as the PM with claude --agent pm, or run scout, engineer, qa or growth directly.
- Let work flow through the queue: the Engineer adds finished pieces to the pending list, QA approves or rejects them with evidence.
- Have Growth upload approved pieces as private only.
- Review each piece yourself and record your yes in the sign-off log before anything is published.
- Add scheduled tasks one agent at a time, and keep any routine that would publish on its own paused.
Common failures
⚠
- A reviewer that judges taste just likes everything. A commenter points out his QA works because each check has a measurable pass or fail.
- Letting the builder review its own work. QA only reports and never edits.
- Two sessions editing the same piece at once; the README uses a per-item lock.
- A commenter flagged that a transcript can't hear background noise cutting out or hiss where a line was removed; measure those with ffmpeg or route them to a human.
- Routines that upload or publish on their own. The README keeps them paused until the owner has reviewed each piece.
Variations
- Not a YouTube channel? The README says the same roles work for most content pipelines: one picks, one builds, one reviews, one packages and grows, and a PM talks to the human.
- One folder per agent with a .claude/settings.json naming that agent, so opening the folder starts the session as it (from his agents README).
- A reply suggests cutting Shorts from a long-form video that already has a corrected SRT and shifting the timings, so captions don't need re-transcribing.
- Use the twelve skills on their own: title optimizer, thumbnail brief, Shorts package, analytics diagnosis, captions and hashtags.
Get the 10 best agent setups. Free.
Plus new recipes every week. One email, no spam.
Unsubscribe anytime. Privacy