50 days ago we posted about analyzing 1.6k Claude Code sessions from our own team. Skills were used in 4% of sessions, 26% were abandoned early, and we had no real benchmark for what good looked like.

Now across 20k+ sessions, we started looking at behavior patterns from derived session metadata: consistency, intensity, session shape, repo breadth, output, cost intensity, and model range.

Nine archetypes fell out, which we turned into playful cards.

We built a Spotify Wrapped meets trading cards experience on top of the classifier: upload your sessions and get a story about how you use AI coding agents.

The classifier runs on metadata like duration, token counts, model mix, repo count, and commit rates.

Create yours today. It’s free and open source: https://app.rudel.ai/wrapped

Repo: https://github.com/obsessiondb/rudel

Comments URL: https://news.ycombinator.com/item?id=48009166

Points: 5

# Comments: 0

I would say roughly equal amount of sessions between them (very roughly)

Also maybe 40% of coding sessions in large brownfield project. 50% greenfield, and remaining 10% non coding tasks.

Or you can run your own instance, but we will need to add docs, on how to control the endpoint properly in the CLI.