I Was Describing My Voice Instead of Showing It

When you set up PodGlue, one of the things it asks is how you speak. There's a box for your tone, another for your style. I filled mine in ages ago with something like "conversational but direct, no fluff, gets to the point."

Then every opening script it wrote for me sounded like a stranger doing an impression of me.

I kept blaming the model. I'd regenerate. I'd tweak the wording in that box, "warm but not chatty," "plainspoken." Small improvements, nothing real. Eventually I stopped reading the generated script at all and just wrote my own from scratch, which rather defeats the point of having it.

The thing I'd got backwards

Last week I went looking at what actually happens when that script gets written, and the problem was sitting right there in the open.

PodGlue was passing the model my description of my voice. Those adjectives, verbatim. That's all it had. There was nowhere in the entire product to give it an actual sample of my writing.

And describing your own voice is genuinely one of the hardest things to do. Try it. Write three sentences about how you talk. Whatever you come up with will be both true and useless, "conversational" describes about eighty percent of podcasters, and it gives a model nothing to hold on to.

Meanwhile I had a folder full of scripts I'd rewritten by hand. Hundreds of paragraphs of exactly how I sound, sitting there, never once used to teach it anything.

I'd been writing a review of my own voice when I could have just handed it a recording.

What changed

The opening script is editable now. That's the small half.

The bigger half is what happens when you save. There's a checkbox: use this as a voice sample. Tick it, and the words you just wrote get kept. Next time PodGlue writes you a script, your real paragraphs go to the model as examples, not "conversational but direct," but the actual thing.

It's told, in plain terms, that where your examples and your description disagree, the examples win. That instruction matters more than it looks. Without it a model splits the difference and you get something that sounds like neither.

You can also leave the box unticked. Sometimes you rewrite a script a particular way because that specific episode needed it, not because that's how you always sound. The edit still saves. It just doesn't get treated as a lesson.

Two things I found on the way

The script was being cut off at four paragraphs on screen. It had been that way for a while and nobody noticed, because the Copy button always took the whole thing, so if you copied it into a teleprompter, it was fine. But the moment I made the text editable, that quietly became a way to delete your own work: open the editor, see four paragraphs, hit save, lose everything after them. Fixed before it could ever happen.

And the obvious way to save it wouldn't have saved it. There's a helper in PodGlue that updates an episode, and it filters what it writes to a specific list of fields, a list that doesn't include the one the script lives in. It would have shown you a success message and thrown the text away on the next reload. That's the second time in a fortnight I've hit that exact shape of bug: something that looks saved and isn't.

What I'd tell you to do with it

One sample won't transform anything. Edit a few scripts over a few episodes, tick the box on the ones that genuinely sound like you, and let it accumulate. The whole thing works on the same principle as everything else here: the useful signal isn't what you say about your work, it's the work.

I spent a year tuning adjectives in a settings box. The fix was showing it a paragraph.