A continuous educational-video pipeline operating a public channel.

Problem

3ED Productions is a content studio focused on educational video. The problem was not quality. It was volume. Producing a single video required four to five roles and one to two weeks of human effort. Growth was linear in headcount, which is to say growth was not growth.

Existing tools could write a script. They could generate images. None of them constituted a pipeline. Each one solved a single step; the team still owned every handoff.

First principles

A video is a sequence of bounded decisions. Topic, hook, sentence-by-sentence script, voice timing, frame composition, thumbnail. Each decision is small. Each can be scored against an explicit criterion. Each failure is cheap to retry.

The bottleneck is not the quality of any individual step. It is the throughput cost of a human in the loop at every step. Removing the human from steps with cheap-retry failures unlocks the throughput. The human stays at the gates where retries are expensive or where taste is structurally human.

Pipeline

The operator triggers a new production from Telegram. The pipeline runs to completion and returns the cut for review.

The motion track runs Seedance 2.0 plus GPT Image 2 with frame-chaining. Gemini 3 Pro runs as the video judge. Identity drift that killed earlier face-based attempts is avoided by the chaining contract: chunk N's end frame becomes chunk N+1's start frame.

Numbers

Proof

The channel is public. The breakout video sits at the top of the iteration arc the pipeline ran through.

The One AI Skill That Beats High IQ People

2,400+ views 64 likes 7 comments No paid promotion

What viewers wrote

Observation

Two patterns surfaced.

Quality gates have to be cheap to retry. A frame that misses the visual standard is regenerated, not surfaced for human review. The cost of an extra generation pass is small. The cost of a human-in-the-loop check at every stage is the original bottleneck. The pipeline retries against its own gates until the gate passes.

The pipeline communicates its reasoning. Output is returned with the decisions that produced it: title scoring, retention markers placed, regenerated frames. The operator overrides at the reasoning level, not the artefact level. The role shifts from reviewer to director.