Rudel: What 1,573 Claude Code Sessions Reveal About How We Use Agents
Rudel is an open-source analytics dashboard for Claude Code and Codex sessions. A CLI hook automatically uploads session transcripts to ClickHouse for processing. Then you get a dashboard showing token usage, session duration, activity patterns, and model usage across your entire team.
But the interesting part is what they found by dogfooding it. From 1,573 real Claude Code sessions with 15 million tokens and 270,000 interactions: skills are only used in 4 percent of sessions. 26 percent of sessions are abandoned, most within the first 60 seconds. Session success rates vary dramatically by task type, with documentation scoring highest and refactoring scoring lowest. Error cascade patterns in the first two minutes predict abandonment with reasonable accuracy.
That 4 percent skills usage number is brutal. The ecosystem has invested enormous effort into agent skills and tools, and real-world data shows almost nobody uses them. The 26 percent abandonment rate, mostly within 60 seconds, suggests that the first impression of an agent session determines everything. If the agent does not immediately demonstrate it understands the task, users bail.
Rudel is MIT-licensed on GitHub at obsessiondb/rudel. Privacy caveat: it ingests full session data which may contain source code, prompts, and secrets, so only enable it on projects where you are comfortable uploading that data. On Product Hunt today.
https://rudel.ai/
← Back to all articles
But the interesting part is what they found by dogfooding it. From 1,573 real Claude Code sessions with 15 million tokens and 270,000 interactions: skills are only used in 4 percent of sessions. 26 percent of sessions are abandoned, most within the first 60 seconds. Session success rates vary dramatically by task type, with documentation scoring highest and refactoring scoring lowest. Error cascade patterns in the first two minutes predict abandonment with reasonable accuracy.
That 4 percent skills usage number is brutal. The ecosystem has invested enormous effort into agent skills and tools, and real-world data shows almost nobody uses them. The 26 percent abandonment rate, mostly within 60 seconds, suggests that the first impression of an agent session determines everything. If the agent does not immediately demonstrate it understands the task, users bail.
Rudel is MIT-licensed on GitHub at obsessiondb/rudel. Privacy caveat: it ingests full session data which may contain source code, prompts, and secrets, so only enable it on projects where you are comfortable uploading that data. On Product Hunt today.
https://rudel.ai/
Comments