Until now, it was basically a cottage industry

One of the types of video content produced and published by ShortWave.STUDIO is something called “OPTOg”.

We used to focus mainly on talk-show-style content under the name “Ping”, but eventually we started making things that no longer fit comfortably within that format, which ultimately led to the name change.

But that is beside the point.

Until now, the video content we produced was largely made by hand.

The workflow looked roughly like this:

1. Write the script in Japanese
2. Use AI to translate it into English
3. Use the script as the basis for creating the main video and audio in HeyGen
4. Use AI for some supplementary assets, such as thumbnail images
5. Open the video editing software, import the assets, and edit everything
6. Export the finished video and upload it to YouTube, the website, and elsewhere

…In other words, a perfectly ordinary workflow.

AI was already involved in areas such as translation and asset creation, but most of the production process itself was still handmade.

A proper cottage industry.

The time was right

At first, that was perfectly fine. Making the content itself is enjoyable, after all.

However, the amount of time and effort required for each video was fairly substantial, and as time passed and circumstances changed, the psychological barrier to sitting down and creating something gradually became higher.

Fortunately, the production workflow itself had become fairly standardised.

The assets we used, the track structure, the settings we applied, and so on had already been turned into templates.

That meant there was a good chance that, if we could work out and verify how to automate what had previously been done by hand, we could significantly reduce the amount of labour involved.

The timing also happened to be favourable.

We now live in an age where tools such as OpenAI’s GPT-6 Astra and Anthropic’s Fable 5.1 are available.

Detailed research, investigation, and even implementation based on those findings could largely be handed over.

So we finally decided to begin the modernisation work we had wanted to tackle for a long time but kept putting off.

Starting with the basics

The first step was to define the policy and organise the requirements.

In other words, requirements definition.

The first important principle was that we would not aim for full automation.

Automation would be limited to “work”, while anything that could be considered “creative” would remain the responsibility of a human.

These two points were defined as principles that must never change.

From there, we would automate as much as reasonably possible.

However, we would compare the effort required with the benefit gained and only implement automation where the cost-effectiveness made sense.

There are plenty of smaller details, but that is the basic idea.

We explained both this policy and the actual production process to AI and asked it to propose several possible approaches.

The eventual direction was to combine MCP, UXP, and direct screen interaction.

Implementation and testing

The main production assets are created using a service called HeyGen.

Conveniently, HeyGen provides an MCP.

After testing it, we found that it covered most of the operations we had previously performed by hand, which meant that nearly the entire process on that side could potentially be automated.

Premiere Pro, however, does not provide a mechanism for AI to automate its operation directly.

To automate the editing process, we needed something that could act as an intermediary.

After investigating several options, we decided to introduce an application called “Premiere Pro MCP”.

https://premiere-pro-mcp.com

As the name suggests, it uses MCP to act as a bridge between AI and Premiere Pro.

There are quite a few products of this kind available now, so if the idea interests you, it may be worth looking into them.

That said, it was not perfect either.

As we continued investigating, we discovered that there were still quite a few things it could not do.

In the end, we settled on a hybrid approach that combines a custom UXP implementation with direct browser interaction through Codex and other methods to automate the workflow as a whole.

What can and cannot be automated

After various twists and turns, we ended up implementing two separate automation skills.

The workflow now looks like this:

1. Write the script in Japanese
2. Use AI to translate it into English
3. Skill #1 reads the script, generates the video and audio, and downloads the results automatically
4. Use AI for some supplementary assets, such as thumbnail images
5. Skill #2 launches the video editing software and processes and places the assets according to predefined rules
6. Take over from the automated process once the initial placement and processing are complete, then make the detailed adjustments manually
7. Export the finished video and upload it to YouTube, the website, and elsewhere

Steps 3 and 5 are the newly automated parts.

These were exactly the sort of tasks that became tedious when performed manually, so automating them has made the overall process considerably easier.

For reference, the division of responsibility across the workflow now looks roughly like this.

What AI and automation handle

• Generating video and audio in HeyGen
• Downloading and organising generated assets
• Checking correspondence between subtitles and dialogue
• Extracting candidate scenes
• Importing and placing assets in Premiere
• Applying predefined initial processing
• Other repetitive tasks that can be expressed as procedures

What humans handle

• Coming up with the concept
• Writing the script
• Overall direction
• Detailed settings and effects for individual assets
• Subtitle adjustment
• Reviewing and correcting AI-generated results
• Deciding when the work is finished

“We chose not to automate it” does not mean “we could not automate it”

A fair amount of manual work still remains in this automation project.

That was intentional.

It is the result of evaluating the cost-effectiveness mentioned earlier.

For example:

• It could be automated, but doing it manually is faster
• It requires case-by-case judgement
• The results are not stable enough
• It consumes too many AI credits
• It involves too much ambiguity
• It is technically possible, but implementing it would cost too much
• It was simply something I wanted to do myself

These are the sorts of reasons why certain parts were deliberately left outside the automation system.

If we reorganised the structure of the videos themselves, we could probably extend the automation even further.

In fact, full automation would probably be possible.

But if we went that far…

It would no longer be ShortWave.STUDIO.

A system that continues to grow

This system has only just been launched, and it may not be perfect yet.

But tools for internal efficiency like this are the sort of thing that should improve through repeated use.

As we continue producing content, we will keep modifying the system based on what we notice and learn along the way.

The aim is to keep building a more efficient production environment over time.