A self-hosted pipeline that researches, scripts, produces, publishes, and learns — one faceless YouTube Short at a time, three times a day, with no human in the loop. It plans a topic, validates it against real competitor data before spending a minute of compute, hand-builds an animated explainer video, uploads it to YouTube, then reads its own analytics every week and rewrites its own playbook to get more views.

It is driven by the Claude Code CLI running headlessly on a Windows box + a local GPU, orchestrated by Windows Task Scheduler, publishing through the YouTube Data API (with Postiz as a scheduling backend).

This is a real, running personal project — not a polished product. It contains one operator's actual channel numbers, ledgers, and strategy notes (kept intentionally, as a build log). Paths, credentials, and the GPU/Docker setup are specific to the author's machine. Treat it as a reference architecture and a build-in-public artifact, not a one-click install. See Honest status below.

                 ┌─────────────── every 12:00 / 17:00 / 21:00 ───────────────┐
                 │  DailyShorts:  pick topic → research (YouTube/Reddit/web)  │
                 │  → validate virality GATE → write script → generate voice  │
                 │  → render animated video → QA watch-back → upload to YouTube│
                 └────────────────────────────────────────────────────────────┘
   weekly ► ShortsLearn:   read our analytics + study top creators → rewrite the rules
   daily  ► CreatorStudy:  frame-by-frame study of a viral competitor → technique catalog
   daily  ► ShortsDigest:  email summary of what shipped + how it's performing
   always ► Dashboard:     old-school live control panel at http://localhost:8899

Full system inventory: future_plans/PLATFORM_MAP.md.

Three ideas shape the whole system:

1. Validate before you produce. Every topic passes a virality GATE (real competitor view-counts
   • a simulated swipe-test) before any compute is spent. Losers are killed on paper.
2. Never look repetitive. YouTube terminates "inauthentic / repetitious" channels. A variation engine forces every new video to differ from the last six on ≥3 axes (title shape, format, visual skin, voice, length, topic cluster) — and every video is a hand-coded animation, not a template fill.
3. Compound. The channel isn't the business — it's the funnel. A weekly learning loop turns real analytics into better rules, and a full monetization roadmap lives in future_plans/.

├── daily_shorts_prompt.md      # the producer brief the AI follows to make + publish one Short
├── daily_shorts.bat            # runner: ensure Postiz up, then run the producer headlessly
├── post_to_youtube.py          # reliable direct upload via the YouTube Data API
├── post_to_postiz.py* / *.sh   # Postiz publish path + OAuth/token plumbing
├── ensure_postiz.ps1           # idempotent Postiz health-check / self-heal
│
├── learn_and_improve_prompt.md # weekly self-improvement analyst brief
├── yt_analytics.py             # pulls channel + per-video CTR / retention into a report
├── yt_top.py                   # finds real top competitor Shorts by view count
├── creator_study_prompt.md     # daily competitor-study brief
├── competitor_playbook.md      # harvested techniques (grows over time)
├── learnings.md                # append-only lab notebook of experiments
│
├── variation/                  # the anti-repetition engine (title shapes, formats, skins, voices)
├── daily_topics.md             # topic backlog with niche/CPM/angle/guardrail
├── daily_posts_ledger.md       # append-only record of everything published
│
├── dashboard.py                # the live control-panel dashboard (stdlib http.server)
├── notify_email.py             # Gmail-SMTP notifier (digest + failure alerts)
├── daily_digest.py             # builds the daily email digest
│
├── .claude/skills/             # the Claude Code skills (ultimate-short, punchy-recut)
├── ClipPilot/                  # the larger video-generation engine (Python pkg + Remotion source)
├── analytics/                  # real performance snapshots (kept as a build log)
└── future_plans/               # the researched monetization roadmap & risk register

*.bat / *.ps1 = Windows runners · *.sh = Git-Bash helpers · *_prompt.md = briefs an AI agent follows autonomously.

The core automation (producer, publisher, analytics, dashboard, learning, email) is Python-stdlib-only. The heavy dependencies (torch, diffusers, edge-tts, faster-whisper, Remotion/Node) are only needed by the video/voice generation layer. See requirements.txt.

No secrets are stored in this repo. By design, every script reads credentials at runtime from outside the repo:

Copy .env.example and follow SETUP.md to wire your own. Never commit real keys. The included .gitignore blocks the common secret files as a second line of defense.

• Orchestration: Claude Code CLI (headless claude -p … --dangerously-skip-permissions) + Windows Task Scheduler
• Publishing: YouTube Data API v3 (direct upload) · Postiz (self-hosted, Dockerized)
• Video: Remotion (hand-coded SVG/TSX explainers) · FFmpeg · optional Z-Image / Wan2GP on GPU
• Voice & captions: edge-tts (rotating voices) · faster-whisper (word timing)
• Analytics: YouTube Analytics + Data API
• Dashboard/notify: Python stdlib (http.server, smtplib)

As of the last snapshot this is an early, pre-monetization channel — the interesting part is the machine, not the numbers yet. Real, un-sanitized metrics, ledgers, and the strategy roadmap are kept in the repo on purpose (analytics/, daily_posts_ledger.md, learnings.md, future_plans/) so the growth curve is auditable.

Known limitations: hardcoded absolute Windows paths (author's machine), a GPU/Docker/Postiz setup you'd have to reproduce, and a Claude subscription as the only real running cost. This is shared as a reference architecture, not a turnkey install.

• This automates content creation. Follow YouTube's Terms of Service and its policies on automated/AI-assisted and authentic content. The variation engine + hand-built visuals exist to keep output genuinely original — don't defeat that.
• Disclose AI-generated/synthetic media where platforms require it (YouTube's altered-content label, TikTok's AI toggle).
• Nothing here is financial advice. Finance content must be truthful and include required disclosures (e.g. FTC affiliate disclosure).
• Respect the licenses and rate limits of every third-party API and asset you connect.

Built on the shoulders of Postiz, Remotion, edge-tts, faster-whisper, FFmpeg, and Claude Code. Third-party model/asset libraries that this project merely uses are not redistributed here.

MIT — do what you want, no warranty. You are responsible for how you use it.