← mattstockton.com

Experiment

The Listening Index

A search engine over my own AI podcast feed.

I built it for myself. I listen to a lot of podcasts and a week later I usually can't tell you which episode a given idea came from, so now every episode I finish gets transcribed and a model reads it and writes the ideas down.

The filter is me. It keeps what's new or useful to my own work in AI consulting, so the page runs heavy on agents and evaluation and light on what I already knew. I put it up because what one person keeps from a year of listening might be worth something if you're working the same problems. Search by topic, show or speaker, or open the Shows tab.

1,214Takeaways
151Episodes
22Shows
143Speakers
142Hours of audio

A model wrote every card here from the episode transcript, and I haven't checked them all against the audio. It works out who was speaking by reading context, and most episodes have one obvious speaker so it gets those right. Where two people talk on equal footing it can mix them up, so if a card has you saying something you didn't say, email me and I'll fix it.

← More experiments at mattstockton.com

How this was made

These are episodes I actually listened to. Everything after that point is automated: a script finds a transcript, a model reads it and writes the cards, and two jobs the model kept getting wrong ended up in code. The colors below mark who decided what.

  • Written by a person
  • Written by a model
  • Decided in code
  1. 219 Episodes in my library Finished, sitting in Apple Podcasts
  2. 151 Transcripts found From the publisher, from video captions, or from Whisper
  3. 1,707 Takeaways written One model call per episode, with the whole transcript in the prompt
  4. 1,214 Cards on this page Sorted, quote-checked, and cut at a score of 3

Getting a transcript

Every episode needs a transcript, and there are three places to get one. I try them cheapest first.

None of these are clean. Automatic captions mishear names and drop punctuation, and Whisper makes up the occasional word, so the excerpts read a little rough. I didn't tidy them up.

Across all 151 episodes it comes to about 1,633,119 words.

151 transcripts, by where the text came from

10 88 53
  • 10 from the publisher. A few shows put a transcript right in the feed, and that text is exact.
  • 88 from video captions. A person wrote 13 of them and the rest are automatic.
  • 53 from Whisper. When there's nothing else, I transcribe the audio on my own machine.

Matching an episode to a video

A feed and a video of the same episode don't line up. Feeds insert ads the upload doesn't carry, so a 23 minute episode turns up as 19, and shows retitle freely, listing an interview by topic in one place and by guest in the other.

Two rules stop the code from guessing.

I've found three wrong transcripts so far, all of them by spot-checking the output. Reading the code wouldn't have helped, since it was doing exactly what I told it to.

The same episode, two lengths

In the feed 23 min As a video 19 min

The shaded blocks are ads the feed adds and the upload doesn't have.

Length only counts when it's distinctive

On a daily show where every episode runs 28 minutes, a 28 minute video doesn't tell you anything. The code only trusts length when at most two other episodes of that show sit within 2% of the runtime.

One video serves one episode

Once a video has been claimed, a later run can't hand it to a second episode.

Who writes what on a card

I use Claude Opus 5 at high effort, one call per episode with the whole transcript in the prompt. It picks out what's worth keeping from an hour of conversation and writes the claim and the paragraph under it. It also pulls the excerpt and scores each one from 1 to 5.

I didn't write any of these cards, and I haven't checked all 1,214 against the audio.

What a card is made of

The headline claim

Model

The paragraph under it

Model

The excerpt

Model picks, code verifies

Who said it

Model, usually reading context

The topic tags

Model

Show, episode, date, link

From the feed

Two jobs I moved into code

Both of these started out as instructions in the prompt, and they moved into code after the model got them wrong.

Putting them in order

I asked for them strongest first and one episode came back 5, 3, 4, 4, 2, so the sorting happens in code now.

Checking the quotes

The code checks every excerpt word by word against the transcript it came from and drops anything that doesn't appear. Across 151 episodes it's dropped 2 excerpts.

Most of the work went into the quotes it rejected by mistake. The model tidies as it copies, so the transcript says "close to using up um regularly" and the quote comes back without the "um". That threw out eighteen real quotes across two episodes before I normalized both sides the same way.

The prompt

The prompt is two files. One is the writing rules. The other is a profile of the reader, who in this case is me.

  1. My own writing Everything on this blog, plus my consulting site
  2. Agents read all of it Ranking the subjects I come back to most
  3. A profile of the reader Which I then edited by hand
  4. Into the prompt For all 151 episodes, and every card on this page

Most of it is about what to leave out:

Anything he has heard forty times. He listens to a daily AI news show; the fortieth "context matters more than the model" is not a takeaway.

The cache key is a hash of both prompt files and the model name. Change one line and every episode is marked stale and gets read again the next time I ask for it.

The score, and the cut at 3

The model scores every takeaway from 1 to 5 as it writes it, and this page shows the 1,214 that came back at 3 or better, out of 1,707.

The score also sets the order you see them in. I don't print it on a card, since it's a judgment about what's useful to me and it would read like a rating of the episode.

Every takeaway the model wrote, by the score it gave

51
4307
3906
Scores of 1 and 2 stay off the page
2481
112

What I quote

These are my notes on other people's work. Every card quotes one passage and credits it.

If you make one of these shows and want a card changed or taken down, tell me and I'll take care of it.

What gets quoted

  • 1,213 excerpts, one per card. Each one is credited and links to the show's own episode page.
  • 28 words on average. That's about 2.6% of a transcript, and 7.1% at the most.