15 million tokens and 270,000 interactions across 1,573 Claude Code sessions. That's the dataset behind Rudel.ai, a new open-source analytics tool built by developers who realized they had zero visibility into their own AI coding workflows.
The project started from a simple frustration: the team was using Claude Code daily but couldn't tell which sessions were productive, why some got abandoned, or whether their usage patterns were actually improving. So they built an analytics layer that connects to Claude Code sessions and surfaces usage data.
The headline finding is striking. Skills - Claude Code's reusable command shortcuts that let you save and replay common prompts - were only used in 4% of sessions. That's a feature designed to save time sitting almost entirely unused, which tracks with what most Claude Code users probably suspect about their own habits: they're using the tool, but not necessarily using it well.
Rudel.ai is available on GitHub as an open-source project. It tracks metrics like session length, token consumption, interaction patterns, and feature adoption rates. For teams running Claude Code across multiple developers, this kind of data could help identify where people are spinning their wheels versus making real progress.
The 4% skills figure is the most quotable stat, but the broader point matters more. AI coding tools are still new enough that most people haven't developed deliberate practices around them. We're in the "just type stuff and hope it works" phase. Tools like Rudel.ai represent the next layer: not just using AI to code, but understanding how you use AI to code. That feedback loop is where the real efficiency gains will come from.