I built a pipeline that auto-generates fact-based documentary videos. The hard part wasn't the render — it was keeping it from making things up.
The pipeline
Source queue → script generation → committee gate → render → publish. Stock footage is zero. The screen is filled only with self-designed cards and edge-tts narration (copyright and tone control). The render itself is the same lineage as the shorts pipeline built on edge-tts + ffmpeg.
The gate is the point
Hand an LLM a fact documentary and its default is "fill in plausibly." So I put a committee stage between generation and publish:
- No ungrounded generation. Numbers, quotes, and causal claims not in the source get cut from the script. If it's not in the source, the sentence is deleted, not inserted.
- Defamation gate. When covering real people and the rise-and-fall of real companies, block flat assertions and smears.
- Investment-advice gate. Since the material is company post-mortems, catch any "buy/sell this stock" that leaks in.
Why this order
Render is cheap and reversible. A published false fact is expensive to reverse and it costs channel trust. So the expensive verification goes before the render — a wrong script never becomes a video in the first place.
A vertical short teaser per long-form
Each long-form auto-generates a vertical short teaser. Shorts don't count toward revenue (YPP watch hours) but I use them as a subscription lever.
The honest part: the value of this pipeline isn't "makes them fast," it's "doesn't make wrong ones." Just as production was never the bottleneck for automation, the real risk in fact content isn't speed — it's hallucination.