Automation · Media
Automated media production chain
A pipeline that produces narrated videos, from script to edit, stopping at every step for me to approve.
In short
- Problem
- Producing narrated videos regularly means going through about ten jobs: script, characters, images, voice, editing, subtitles. None is hard, but repeating them every week quickly becomes unsustainable.
- Solution
- A 14-step pipeline, from script to a video ready to publish, with my approval between each step and a maximum budget per episode.
- What it brings
- Production becomes regular and its cost is known in advance, without letting the machine decide on the content.
- Where it stands
- Ongoing. The pipeline produces complete episodes, but publishing is still done by hand.
- My role
- I split the pipeline into steps, defined the approvals and the budget, and steered its build with coding assistants.
The project
I wanted to produce narrated videos regularly: La Fontaine's fables, and a story set in ancient Rome. Each video takes about ten different jobs: writing the script, keeping characters consistent, creating images, voice, editing, subtitles, delivery.
No single step is hard. Repeating them every week is what does not hold up.
Why not one program that does it all
The first reflex is a single program that goes from topic to finished video. It does not work, for three reasons.
When it breaks, you lose everything. One error halfway, on an image, and all the earlier work has to be redone.
You find out the cost afterwards. Generated images and voices are paid per use. Without a limit, you learn what an episode cost once it is finished.
Automating everything gives a worse result. A pipeline that picks its own topics and publishes without review produces exactly the kind of content platforms push down.
What I built
A 14-step pipeline, each step independent, with an approval between each one.
- A step can fail without breaking everything. Progress is kept, and you pick up where it stopped.
- The budget is set upfront. If the cost goes over the cap, the pipeline stops rather than carrying on.
- Every step waits for my go-ahead. AI does the production, not the choices.
What the pipeline does not do
It does not find its own topics or fact-check: it starts from a topic I give it. And it does not publish on its own: it prepares a complete episode, ready to go, which I put online myself.
Show technical details
The steps
They cover the script, character consistency, images, voice, editing, music, checks, delivery, publishing, progress tracking, a report and a dashboard. Progress tracking is the central piece: it is what makes it possible to resume at the step that failed.
Editing
Everything goes through FFmpeg: still images animated with a camera movement, subtitles, mixing voice and music, adjusting the volume.
Short formats
A separate tool produces vertical clips in 1080 × 1920, with subtitles generated by automatic transcription and volume adjusted for social media.
Budget
The cap is read at start-up. Every paid generation is counted, and the pipeline stops cleanly, keeping its progress, if the cap is reached.
Numbers
14
steps in the pipeline
script, characters, images, voice, editing, checks, delivery, publishing, progress tracking and dashboard
Source · inventory of the project's pipeline directory
7
episodes ready to publish
a complete series of fables, produced end to end
Source · ready-to-publish episodes directory
3
series
three worlds, each with its own visual style and characters
Source · project episodes directory
What I did
Architecture
I split production into 14 independent steps, so a failing step does not bring the others down.
Needs definition
I decided what should be automated (production) and what should not (choosing the content).
Assisted development
The pipeline was written in Python with coding assistants, using FFmpeg for editing and automatic transcription for subtitles.
Quality control
I set an approval between every step and a maximum budget per episode.
Documentation
I documented how the pipeline works, the tools it uses and its limits.
Evidence
A 14-step pipeline
verifiedEach step is independent and keeps track of its progress. A step can be rerun without restarting everything, and a failure does not wipe out the work already done.
- Source
- project pipeline directory
- Verify with
- Listed the steps and read the progress tracking.
Seven episodes produced
verifiedA complete series was produced by the pipeline, from script to an edited, subtitled video ready to publish.
- Source
- ready-to-publish episodes directory
- Verify with
- Listed the episodes on disk.
A maximum budget per episode
verifiedThe pipeline stops if the cost of generation goes over a cap set in advance. You know what an episode costs before launching it, not after.
- Source
- chain configuration
- Verify with
- Reading the configuration and the budget control mechanism.
Short formats
verifiedA separate tool turns a video into a vertical clip for social media: cutting, automatic subtitles and adjusted volume.
- Source
- the project's vertical formatting tool
- Verify with
- Read the tool and its settings.
Limitations
- Publishing is not automatic: the pipeline prepares an episode ready to go, and I put it online myself.
- The pipeline does not find its own topics and does not fact-check. It starts from a topic I give it.
- Choosing the topic, the hook and the angle stays with me, on purpose: that is where everything is decided.
- It depends on external services for images and voices. If their prices or how they work change, part of it has to be redone.
- Platforms are increasingly limiting mass-published content. One more reason to keep a human approval.
What I take away
- The reflex is to want to automate everything. The real gain comes from automating repetitive production and keeping the content choices, where the difference is made.
- Splitting into independent steps changes everything: when a single-block pipeline breaks halfway, you lose it all. With steps, you pick up where it broke.
- A hard budget set before launching beats a cost-tracking dashboard you look at afterwards.
What it could bring to a company
- A marketing team can organise its content production the same way: assisted generation, approval at every step, controlled cost.
- A training team can produce internal videos from approved scripts, with automatic subtitles.
- Any repetitive multi-step task (reports, product sheets, translations) can be split the same way, to resume at the step that failed.
Technologies
- Python
- FFmpeg
- Whisper
- Bash
- Image and voice generation APIs