Reels Engine

An AI clone of you, posting daily. This engine produced 931K
impressions in 90 days for ~$34 in render credits —
watch the full breakdown of exactly how.

Automated short-form pipeline:
trend research (GitHub / Hacker News / RSS) → LLM-written scripts in YOUR
voice → HeyGen avatar clone → Remotion graphics + karaoke captions → a QA
gate that checks every frame (face position, audio loudness, black frames,
caption accuracy) → scheduled to every platform via Metricool. Plus a
2-hourly breakout watcher for big trends and a metrics-driven style memo
that makes every script better than the last. You approve each video with
one tap in Telegram. Nothing posts itself.

Built and run daily by Johann (@johannships) —
this is the actual engine behind the channel, not a demo. I share how I
operate it profitably (my live configs, funnel numbers, what the metrics
taught it) inside AI Operators.

Try it in 60 seconds (no accounts, no keys)

Only needs Node 20+. Renders a real episode's graphics so you can see what
the machine makes before you sign up for anything:

git clone https://github.com/johannships/reels-engine && cd reels-engine
mkdir -p episodes/demo-episode && cp examples/demo-episode/* episodes/demo-episode/
cd studio && npm install && npx remotion render RepoRadar \
  ../episodes/demo-episode/canvas.mp4 --props=../examples/demo-episode/props.preview.json

Everything past this point is accounts and keys (HeyGen, Metricool,
Telegram) — Python 3.11+, ffmpeg, and whisper.cpp only enter at the full
pipeline stage, and the whisper model auto-downloads on first run.

Set it up with Claude Code (recommended)

git clone https://github.com/johannships/reels-engine
cd reels-engine && claude
> set this up for me

CLAUDE.md teaches the agent the entire onboarding: a niche interview
that personalizes the writing voice, creating your HeyGen clone from zero
(including how to record the training clip so your clone doesn't fidget),
wiring Metricool + Telegram, a local proof run, and deploying to Railway,
a VPS, or a Mac mini. First reel in about an hour, most of it waiting on
renders. Prefer manual? CLAUDE.md reads fine as a human runbook — copy
pipeline/.env.example → pipeline/.env and go stage by stage.

What it costs to run

• HeyGen API: ~$1 per minute of rendered avatar (the only real cost)
• LLM: ~free on a flat-rate coding plan (any Anthropic-compatible endpoint)
• Metricool Advanced (posting API) · Railway ~$5-10/mo or your own box

How it works

research.py / topics.py / watch.py     what's trending in YOUR niche
        │
scriptgen.py + script-skill.md         hooks, beats, payoff discipline —
        │                              craft bible + ruthless-editor LLM
        │                              pass, tuned by your own metrics
heygen.py                              your clone, one continuous take
        │                              (looks rotate via avatarLooks)
prep.py                                whisper word timestamps → Remotion
        │                              graphics (repo cards, screenshots,
        │                              kinetic text) → ffmpeg composite
qa.py                                  format/loudness/black/freeze/
        │                              captions/face-headroom, self-
        │                              correcting crop loop
approvals.py + telegram                YOU tap approve
        │
metricool.py                           scheduled everywhere at once

pipeline/   the python pipeline · config.json (neutral defaults) ·
            config.local.json (YOUR identity — gitignored, see configlib.py)
            style-memo.md (the system's memory) · .env (secrets, untracked)
studio/     Remotion template (1080x1920, split/fullscreen layouts)
deploy/     Railway · VPS · Mac mini paths
examples/   a committed sample episode you can render without credentials
episodes/   (untracked) one folder per reel: research → script → final.mp4

Your identity never lives in the repo: config.local.json holds your
name, voice rules, niche keywords, avatar looks, and funnel words,
deep-merged over neutral defaults.

Hard-won architecture notes:

• Avatar video is composited by ffmpeg, never played inside Remotion
  (headless Chrome crashes; see prep.py).
• Fonts are CSS @font-face data URIs (FontFace API + delayRender hangs on
  Remotion's page recycling).
• Caption/scene alignment matches each scene's first two words against the
  Whisper transcript (list videos keep First/Second/Third openers). Scene
  openings must stay distinct — and transition slop ("Now the setup") is
  banned outright in pipeline/script-skill.md §3a.
• Whisper mishearings self-correct against the script's own vocabulary,
  so your CTA keyword always renders exactly (see fix_misheard in prep.py).

Docs

Operating manual: CLAUDE.md (setup) · PIPELINE.md (daily ops) ·
OPERATOR-HUMAN.md (your 10-min/day job) · OPERATOR.md (24/7 agent
guardrails) · CLONING.md (instances for teammates/clients) ·
FUNNEL.md (comment keyword → DM → email → community monetization)

Author's context notes (how this system came to be — interesting, not
required for setup): GOLDIE-NOTES.md · HERMES.md · CYNDRA-AGENT.md

License

MIT. Contributions: see CONTRIBUTING.md — this is my live daily driver,
so I merge conservatively.

If this replaces an editor for you, star the repo — it's how other
builders find it.