shorts-autopilot

Autonomous daily YouTube Shorts + Instagram content pipeline for the MyProposer brand — topic in, published Short out, no hands on keyboard.

Interactive, pannable version: open the architecture map

What it is

A Python + Remotion pipeline that runs on a schedule and publishes finished vertical videos for MyProposer (my AI proposal-software product). One run takes a topic through script generation, voiceover, music, render, mux, and a headless YouTube Data API upload with a real custom thumbnail. A companion job reposts the best-performing Short to Instagram Reels. Status: automated and posting on schedule — YouTube Shorts autopilot live, Instagram repost automation live (first Reel published July 2026).

Why I built it

MyProposer needed a steady top-of-funnel content presence, and I make videos for a living — but hand-producing 2–3 Shorts a day is a full-time job that competes with building the actual product. So I built the producer: a pipeline that writes, voices, scores, cuts, renders, and publishes on its own, and only asks for a human when a step genuinely can’t run headless.

Architecture

cron (3x/day) -> schedule guard -> run_short.py
  topic + style rotation
    -> staged story director (5x headless `claude -p`)  [AI-image story path]
       or style writer (motion / editorial / punchy)
    -> ElevenLabs TTS (char-aligned timing) + music bed
    -> Remotion render (AutoShort | EditorialJourney | PunchyType | PhotoStory)
    -> ffmpeg sidechain-ducked mux, -14 LUFS
    -> YouTube Data API upload + custom thumbnail
run_instagram.py -> performance ranker -> Upload-Post API -> Instagram Reels

Layer	What happens
Entry points	`run_short.py`, `run_story.py`, `run_instagram.py`, `run_scheduled.py`, cron wrapper
Scheduling	Fixed cron entries + a config-file schedule guard (pause/resume/slots without touching crontab)
Creative	5-stage story director over headless Claude Code CLI, plus per-style script writers
Audio	ElevenLabs TTS with character-level alignment, ElevenLabs Music bed (graceful degrade)
Render	Remotion compositions sized from measured voiceover duration; ffmpeg ducked mux
Publish	YouTube Data API v3 (resumable, cached refresh token), thumbnail set separately
Distribution	Views + 20×likes ranker picks the repost; Upload-Post API posts to Instagram
State	Per-run `output/<run-id>/` (resumable stages), FIFO story queue, small JSON state files

Every stage writes an artifact and is skipped on rerun, so a failed run resumes with the same run-id instead of re-spending TTS quota.

Engineering highlights

Headless LLM without an API key. Script generation shells out to claude -p on my Claude Code subscription with a durable setup-token, because an interactive login doesn’t survive non-interactive subprocess use. Five staged calls (premise → greenlight → script → shot parts) replace one, because a single prompt asked to invent a story, meter it to 15 cuts, and write cinematography optimizes three conflicting losses and fails at all three.
An independent greenlight critic. Stage B is a fresh Claude call given no product docs on purpose — its job is to reject any premise whose turn is “they found a tool.” Premises are shuffled before judging so position can’t bias the pick.
Character-bible anti-drift system. The LLM never writes a finished image prompt. It locks a character bible once; Python composes every shot prompt around that identical wording — built after confirmed drift between shots (grey hoodie in one frame, different wardrobe the next). Validators enforce TEXT_BUDGET = 0 (generated lettering once produced “BARNNISE COILECTION”; product text lives in the crisp Remotion caption layer instead), a face budget of 6/15 shots, and a timeline gap/overlap check.
Measured, not assumed, pacing. The words-per-second constant is 2.16 — measured from a real 77-word ElevenLabs read at 35.68s. The original 2.7 estimate under-predicted duration by ~25% and pushed every cut past its cap. Scene timing itself comes from ElevenLabs character-level alignment, not LLM guesses.
Quota-aware error surfacing. ElevenLabs reports an exhausted character quota as a bare HTTP 401, which reads like an auth failure. The pipeline inspects the error body and raises a QuotaExceeded with the real diagnosis — “top up the plan; every scheduled post is blocked” — instead of sending me chasing a phantom key rotation.
The silent story starvation post-mortem. The AI-image story slot posts from a queue (image generation is session-only, so stories are batched and queued). When the queue ran dry, a fixed fallback quietly gave one style 2 of every 3 posts and made the channel look repetitive. Now an on-disk starvation counter makes every degraded slot loud, and the fallback style rotates so no single look dominates.
Upload guardrails. Headless YouTube Data API upload with a wrong-channel guard — if the video lands on any channel other than the configured brand channel, the run fails with delete-and-reauthorize instructions instead of silently posting to the wrong account. Thumbnails are uploaded via the API (kept under YouTube’s 2MB cap) rather than baked in as a cover frame, so Shorts play the instant they’re tapped.
Scheduling around macOS TCC. macOS blocks crontab writes from the agent shell, so cadence is enforced by a config-file schedule guard read at runtime — pause, resume dates, and posting slots change by editing JSON, never crontab (launchd as the wake-proof fallback). Sidechain-compressed ffmpeg mux keeps music auto-ducked under the voice, normalized to YouTube’s -14 LUFS.

Stack

Layer	Tech
Pipeline	Python 3 (`content_pipeline/` package)
Script generation	Claude Code CLI, headless `claude -p`
Voice + music	ElevenLabs TTS (with timestamps) + ElevenLabs Music
Video	Remotion (React/TypeScript compositions), 1080×1920
Audio finishing	ffmpeg (`sidechaincompress`, `loudnorm`)
Publishing	YouTube Data API v3 (OAuth, resumable uploads)
Instagram	Upload-Post API
Scheduling	cron + config-file guard, launchd fallback

Status & roadmap

Running: scheduled Shorts posting with custom thumbnails, and Instagram reposts ranked by engagement. The AI-image story format is the primary creative direction; keeping its queue filled is the main operational task (image generation still requires an interactive session). Next: automating queue refill further and widening the topic/deck rotation.

About this repo

This is a public architecture showcase of a private production codebase. The source stays private; this repo documents how it’s built.

— Godfred Aidoo · godfredaidoo.com · LinkedIn · more projects