An automated video production system
A topic goes into a database table. A script gets written from source material, narrated sentence by sentence, matched to vetted archive imagery, rendered with burned-in captions, and published on a schedule. No camera, no microphone, no editor.
Durations are measured on a four-core VPS, not estimated.
A research pass proposes topics and fetches the source text each one will be written from. Nothing enters the queue without material behind it.
~25 topics/day
The queue is served against a rolling 24-hour cap, not a calendar day — with an in-flight claim, a three-strike rule, and a duplicate check that compares the core of a title rather than the string.
4 clauses, 4 outages behind them
A language model drafts from the source text and is held to a length, sentence count and opening structure. Failures are rejected and retried rather than shipped.
165–195 words
Each sentence is synthesised separately so its start and end are measured, not guessed. Captions cannot drift, because nothing about their timing is estimated.
1 API call per sentence
Images come from the topic's own encyclopaedia article and its linked articles first, then from licensed archive search — and every search result is checked by a vision model before it can appear.
24 images/chapter, 13 from the archive
One clip per image, drifting slowly, cut on an irregular rhythm so it reads as a film rather than a slideshow. Captions are burned in; loudness is normalised once across the finished audio.
21 s of CPU per 5 s of video
Resumable upload, custom thumbnail, playlist, and a database row recording every production parameter — so that later you can answer whether a change helped.
recorded, not just uploaded
Anyone can generate code that calls an API. What you cannot generate is knowing where it breaks. Each of these cost days, and each one is in the source with the reasoning attached.
The daily cap counted videos published “today”. At one minute past midnight it reset and pushed a full day on top of videos four hours old. Everything past roughly the thirteenth video in any 24-hour window collapsed to a fraction of the views before it.
Fixed by counting a rolling window instead.
Four API keys appeared exhausted at once. Testing per model showed the 429 came from one model only, while the same keys answered instantly on another. A background job had been draining exactly the model the pipeline fell back to.
Cheap work moved to a cheaper model.
A single ffmpeg call carried a 60-second timeout. Measured on an idle machine a five-second clip takes 21 to 28 seconds; under load it takes longer. Three finished documentaries died at that limit, each after every chapter had already been narrated.
Raised, and the expensive grain filter removed — it cost 38% of render time and survived no compression.
A description containing “the world’s first” ended an SQL string literal and killed the insert. The video was already live on YouTube; the database never heard about it, so the topic looked unused and was queued again.
Parameterised queries, and a rule against string interpolation.
$9
The handbook. Architecture, the database schema, and the fourteen production failures with their fixes.
This is the blueprint, not the building.
Most complete for the money
$29
The whole working framework, generalised for any niche, plus the setup prompt that lets an AI agent install it for you.
Node 20+, ffmpeg, Postgres.
$59
Everything above, plus the video masterclass in English and German.
Buy the Blueprint first? Its price comes off this one.