What it does
Over a week of real use, about one suggestion in twenty was taken. The writeups walk through why and what the author kept anyway.
Agent tooling
A Claude Code plugin that logs which skill it thinks fits each prompt and mostly stays out of the way. The numbers it collected argue against routers like it.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Over a week of real use, about one suggestion in twenty was taken. The writeups walk through why and what the author kept anyway.
Try recording suggestions without showing them to your coding assistant. This is the tool's default setting. Compare the suggested skills, or job instructions, with the ones the assistant actually uses. Your requests and skill descriptions go to an online AI service.
The author recorded suggestions for a week. Only 28 of 539 suggested skills were used within 30 minutes in the chat where they were suggested. Use that result as a reason to measure, not assume the suggestions help.