The Best Moments Of 68 Episodes Were Being Thrown Away

Open an episode in PodGlue and you get three things back from the conversation: the themes you kept circling, the follow-ups you promised, and the key moments, the bits worth clipping, quoting, putting on a slide.

Two of those had been working for months. Themes: 396 of them. Follow-ups: 163.

Key moments: zero. Not a low number. Zero, across every episode we'd ever analysed.

The part that made it hard to spot

All three come from the same place, at the same time. One pass over your transcript, one answer, three lists pulled out of it and saved.

So "themes work but moments don't" made no sense. It's not two features, it's one. If the analysis had failed, all three would be empty. If it succeeded, all three should have something.

And nothing anywhere was broken. No error. No red banner. The screen just said there were no key moments in your episode, which is a perfectly believable thing for a screen to say. You'd look at it and think: fine, maybe that conversation didn't have a standout line.

Sixty-eight episodes in a row. Every one of them had standout lines.

What was actually going on

When PodGlue asks the AI to read your transcript, it asks for specific things by name. For themes it says: give me the theme. For follow-ups: give me the action text. Then it goes looking for exactly those.

For key moments it asked for an excerpt, and then went looking for a transcript excerpt.

That's the whole bug. One thing asked for, a slightly different thing looked for. The AI did its job perfectly every time. It handed back eleven beautiful moments with the quotes attached, and PodGlue looked at each one, couldn't find a field with the exact name it expected, and quietly binned it. Eleven for eleven. Then it reported that it had found nothing, because by that point it genuinely had nothing left.

The two features that worked, worked for the boring reason that somebody had written the field names out properly. The one that failed was the one where we'd described what we wanted in a sentence instead of naming it.

What shipped

I didn't want to guess at this one, so I ran the old version against a real transcript and watched it happen: eleven moments in, zero out. Then the fixed version on the same transcript: seven in, seven out.

Now PodGlue asks for the field by its exact name, tells the AI to hand it back in a fixed shape rather than trusting it to guess, and, this is the part I care about most, accepts the quote under any of the obvious other names it might come back as. Three separate things now have to go wrong for a moment to be lost, instead of one.

It also keeps count out loud. If the AI hands over ten moments and PodGlue saves zero, it now says so, loudly, instead of shrugging and showing you an empty page. That's the actual lesson here. The bug wasn't expensive because it was subtle. It was expensive because "we found nothing" and "we threw everything away" looked identical from the outside.

Here's a real one it saved on the first run after the fix, from an episode about a genetic condition diagnosed in childhood:

I think the biggest obstacle is the alone feeling of it.

That was sitting in the transcript the whole time. Timestamped, quotable, ready to go on a card. It just never made it out of the machine.

What's next

The fix only helped new episodes, so the back catalogue needed a pass of its own. That has now run, across all 68 episodes that had been analysed with nothing to show for it. Carefully, though: re-running the whole analysis would have trampled the follow-ups people have already ticked off, so it re-did the moments and left everything else exactly where it was.

So Key Moments and Quote Suggestions are finally rendering against real episodes, which they had never once done. That first pass was noisy, though: some episodes came back with the same line more than once, and a few excerpts ran long enough to stop being quotable. Both are fixed now, with repeats dropped and a ceiling on how many any one episode can claim. Better eight good ones than fifteen with the same line three times.

If you've been looking at an empty Key Moments panel and assuming your episodes just didn't have a moment worth pulling: they did. We were dropping them on the way out.