Voice
Warm, confident British male. Corporate and cinematic, but not salesy.
Video translation workflow
trying a different route to Smart Cat
A meeting I was only half-listening to became an experiment in video production, translation and AI-assisted building.
Standalone
Smart Cat comparison
Codex workflow
Fictional source asset
German and French next
Greg's take
I wasn't paying much attention during a meeting when somebody mentioned video. That turned into a message, a scripting error, and a question: if Smart Cat couldn't get the translation moving, could I take a different route and build the whole workflow myself?
The slightly awkward answer is yes.
This started as a workaround. It became a comparison. On one side, a process that stopped because a script fell over. On the other, Codex coordinating different AI models, building the assets, refining the video, preparing the translation route, and leaving behind a learning layer that can be reused next time.
That last bit matters. A translation workflow should not just process a file. It should learn from the decisions, fixes and preferences that happen along the way.
Which is why the name Smart Cat is doing quite a lot of heavy lifting here.
Video test asset
A fictional film, built to make the next test possible.
To test the translation workflow safely, I first created a fictional 30-second rebranding film for Smylex, a new oral-care healthcare company separating from ACME.
Safe source asset
A fictional company film avoids handing unreleased material to an AI system.
Translation-ready test
The same video can now be used for German and French translation comparison.
Controlled review
The workflow keeps review, refinements and cost assumptions visible.
The start
The practical problem arrived sideways.
I wasn't really paying attention during a meeting when someone said something about video. Afterwards, I sent them a message to find out what I had missed.
They explained that they were having trouble getting a video translation to work because of a scripting error. I decided to test whether I could take a different route and use Codex, with access to different AI models, to manage the workflow instead.
There was one immediate problem: you cannot simply hand an unreleased company video to an AI system.
So before I could test the translation, I first had to create a completely new fictional video.
That became my first build.
The brief I actually needed
Small enough to start. Specific enough to steer.
Create a credible 30-second rebranding film for Smylex, a fictional oral-care healthcare company becoming independent from ACME.
It should feel premium, warm, serious and suitable for healthcare professionals across Europe, the Middle East and Africa.
Show its established oral-care heritage, modern manufacturing, clinicians and patients, then finish with the Smylex identity.
The only answers that really mattered
A few decisions did most of the work.
Music
Technology-focused, energetic and progressive. It needed to build towards the end, not sound like soft corporate plinking.
Refinements
Make the clinician's hand movement look natural. Make the workshop team move like real people. Keep the outside-to-workshop camera move.
What I used
The build was less magical than it sounds. Good.
Codex was used to plan the workflow, write prompts, organise files, assemble outputs and prepare the final package.
A generative video model was used to create and refine the moving scenes.
A text-to-speech model was used for the English male voiceover, using the Cedar voice with a neutral British corporate direction.
A supplied music file was used as the background track.
The final edit combined the video, voiceover and music into a 30-second film.
The video was refined through direct feedback: smoother people movement on the workshop floor, a restored outside-to-workshop camera move, and more realistic dentist-patient motion.
The same workflow is intended to be used next for German and French translation.
Tomorrow's test
The real point is the translation workflow.
Tomorrow, I'm going to show the team this video translated first into German and then into French. After that, I'll show them what the experiment actually cost.
The comparison
Codex vs Smart Cat
The useful comparison is not "which name sounds cleverer". It is what happens when a real video translation job gets stuck, and whether the route you choose leaves you with something reusable afterwards.
| Decision area | Smart Cat | Codex |
|---|---|---|
| Starting point | Started inside the translation platform. | Started by creating a safe fictional test asset. |
| Failure mode | The test stalled because of a scripting error. | Produced the Smylex video first, then prepared the translation route. |
| Time to first useful output | Time was lost before translation could begin. | Time went into building a reusable workflow, not just pushing one file through. |
| Estimated cost visibility | Cost comparison is not yet visible from the stalled run. | Cost can be estimated from model calls, generation steps and edit passes, then measured properly tomorrow. |
| Control over the workflow | Workflow control depends on the platform process. | Workflow control is visible and adjustable. |
| Model flexibility | Model flexibility appears constrained by the platform route. | Different AI models can be used for video, voice, translation, review and assembly. |
| Refinement | Refinement is mainly inside the translation process. | Refinement can happen at every stage: script, visuals, voice, music, edit and translation. |
| Reusable learning layer | The reusable learning layer is less visible to the user. | The reusable learning layer can live directly in the repo as prompts, preferences, terminology, checks, scripts and notes. |
| Human review | Human review is still required. | Human review is still required. |
| Likely next-step effort | Fix the script before the translation can be tested properly. | Translate the video into German and French, then compare time and cost properly. |
This is the bit I care about. Not whether one file can be forced through one process once, but whether the next file gets easier. With Codex, the workflow can remember the decisions: tone, terminology, voice preference, visual fixes, review checks and cost assumptions. That turns the exercise from a one-off translation attempt into a reusable production route.
Closing
The next file is the real test.
I started with a meeting I was barely following and ended with a film I had built from scratch. The next question is whether this different route can produce the translations, and what the numbers look like when it does.
The interesting part wasn't that the AI produced everything from one sentence. It was how little information it needed to begin, and how quickly each piece could be refined through ordinary, direct feedback.