The AI makes the video. The client makes every decision.

Chrvoje Engineering runs an engineering channel. The request, in the end, was simple to say: I want to type an idea and get a finished video.

That is easy to promise and easy to do badly. A system that turns a sentence into a video with nobody looking is how the internet filled up with content nobody wants to watch. So the pipeline we built is mostly about the opposite question: where does it have to stop?

It starts with one line

The client types a video idea. One line, the way you would say it to a colleague.

That is enough because the system is not starting from nothing. It holds the channel's context: what has been published, how it sounds, who watches it. A general-purpose tool given the same sentence writes a general-purpose script. This one writes the next video for this channel.

Stop one: the script

The system drafts the script and then waits.

The client reads it, changes what is wrong, and approves it. Nothing else happens until they do. This is the cheapest place in the whole process to fix a mistake, because so far the only thing that exists is text. A wrong fact caught here costs a minute. The same fact caught after the video is made costs the video.

Stop two: the whole video, before any of it exists

This is the step people do not expect.

Once the script is approved, the system plans every shot and writes the prompt for each one. Then it shows the client the entire video laid out: every shot, in order, with what each will be.

Nothing has been generated yet. That is the point. Generation is the expensive part, in time and in cost, and a video generated from a plan nobody checked is mostly waste. Seeing the whole structure first means the client is approving a plan, not discovering one.

Then it generates, and everything stays editable

With the plan approved, the system produces the shots and the voice.

Any of them can be changed afterwards. A shot that came out wrong is redone on its own; the voice is handled the same way. The client is not choosing between accepting the whole thing and starting again.

Stop three: the thumbnail

The system makes a custom thumbnail, and the client approves it. It is a small step and it is deliberately not automatic: the thumbnail decides whether the video gets watched at all, so a person looks at it.

Then it publishes

Once the thumbnail is approved, the video is uploaded automatically. The blog post and the article that go with it are published alongside, so one idea becomes the video and the writing around it without a second round of work.

What this is actually automating

Count the stops. Script, plan, thumbnail. Three places where a person decides, and everything between them done by the system.

That ratio is the design. The client still makes every decision that makes the channel theirs. What they no longer do is the repeat work between the decisions, which was most of the hours.

Across formats the channel has passed two million organic views, with no paid promotion behind it.

If you are thinking about a pipeline like this

Give it your context, not just a prompt. The difference between generated content and your content is everything the system knows before you type.

Put the approval before the expensive step. Anything costly to produce should be approved as a plan first.

Keep every piece changeable. A pipeline you can only accept or reject as a whole will be rejected as a whole.