Key Takeaways

  • Recommendation algorithms degrade when multiple people share an account, polluting curated queues with kids' media or irrelevant clicks.
  • OpenAI Product Lead Kath Korevec replaced Spotify's algorithmic discovery by scripting AI computer use to crawl top tracks on Reddit music subreddits every Monday at 8 a.m.
  • The pipeline feeds scraped tracks directly into a custom interface backed by Spotify API connectors and local Apple Music files.
  • Browser-operating agents let you extract structured data from community forums that have locked down traditional API access.

The Method

Recommendation engines treat every click as an endorsement. For Kath Korevec, sharing an account with her children meant her algorithmic feeds collapsed into cartoons and nursery rhymes: “My Spotify has been taken over by my children. And yeah, some of it's Michael Jackson, which is great, but most of it's Bluey and Sesame Street.”

Instead of fighting the algorithm or manually searching for new artists, Korevec built an automated pipeline using ChatGPT Sites, Codex, and computer use.

The workflow runs on three distinct components:

First, Korevec schedules an AI agent to run every Monday at 8 a.m. The agent opens a browser, navigates to Reddit's top music communities, and inspects the most upvoted tracks and discussions from the preceding week. As Korevec explained: “I have it go to Reddit and use computer use to look at what people are sharing. And then I put that in an automation. I have it run every single Monday at 8 a.m. And then it builds a playlist for me.”

Second, the system parses track titles and artist names out of unstructured forum posts, deduplicating submissions and ranking them by community upvotes.

Third, the curated list syncs to a custom playback interface. Host Claire Vo observed: “I didn't realize that Spotify was a plugin that you could use. And so it can query the catalog and pull it in to play in whatever custom interface that you want.” Korevec links this setup to both streaming and local files: “I have it using Spotify for the actual player pulling in the music. But then it can also reference my Apple Music and pull that in because I do have a lot of music on my local machine.”

Where This Breaks Down

Relying on computer vision and browser automation introduces structural fragility that direct APIs avoid. Reddit regularly updates its DOM layout, tests anti-scraping checks, and tweaks front-end UI components. When an interface element shifts, agentic vision models can misclick, drop out of pagination loops, or misread upvote counts.

Community upvotes also optimize for broad consensus rather than niche personal taste. A subreddit top list filters out Bluey, but it replaces personal discovery with the mean preference of thousands of anonymous commenters. If you want high-precision curation, community scraping requires secondary AI filtering against your historical listening logs.

What to Do With This

Pick one repetitive discovery task where a platform algorithm serves you generic or corrupted recommendations. Build an automated prompt that instructs an agent to browse two niche discussion forums every Monday morning, extract the top three mentioned tools or resources, and push them to a private Slack channel or Notion database.