What worked?
I managed to prompt my way to an LLM listening to my voice, cloning it, and adding narration to short videos with subtitles. Videos made it to YouTube and Facebook. That was the exciting part.
Then I asked for a reusable skill and scheduled uploads. In the original LinkedIn post, I said the scheduler had failed for reasons still to be determined. Prompting while I was away from the laptop was inconsistent, or did not produce the result I expected.
I am describing my own experience. I have not established one cause for every failed run, and this is not a controlled comparison of AI products.
Where does a working demo stop?
A finished video answers one question: did the rendering step produce a file? It does not answer whether the right channel received the file, whether the upload stayed private, or whether a scheduled run actually published it.
Those differences matter when I step away expecting work to continue. A useful handoff has to preserve the exact stage reached. If a draft exists, the next session should recover that draft rather than upload another copy.
What do I mean by subtitles?
My cloned voice reads the script. Matching subtitles are uploaded as a separate timed track, so viewers can turn them on or off in the player. I do not want narration subtitles permanently painted into the video.
Designed titles, dates and source labels can remain in the picture. The words being read belong in a subtitle file as text. I want that text checked against the narration, including its timing.
Having a subtitle track does not establish complete accessibility compliance or guarantee that a search engine will index the video. It is one part of making the content usable and available as text.
How do I know what is public?
I need a record with the brand, destination channel or Page, actual public link, subtitle status and source file. An attempted upload and a public video are different states. A processing message is not a publication receipt.
I also want each brand's files in its own repository folders. Videos, scripts, subtitles and receipts should be together. Private reports need protection from accidental website publication. A knowledge-search folder should not become the place where all the deliverables disappear.
What does this have to do with traffic?
The content needs a useful next step. A video about a specific page should point to that page. A readable article gives the same idea a home outside the social feed, where people can read it without watching the video.
Then I need to distinguish people seeing the content, visiting the site, and taking a useful action. They are separate outcomes. Publishing more files does not, by itself, answer any of them.
What I want from the next run
Show me what was created, where it was saved, and where it is live. If something stopped, tell me the exact step and keep enough information to continue without starting over.
I am documenting these lessons while building Prism Publication. You can try the live ad-testing interface or read how I think about connecting useful answers to buyer questions.