Question for the community before I get into the data: how do you currently keep track of what your installed skills are doing on a given call, and how do you decide which ones to remove? Asking because I tried to answer both questions for myself last month and the answers were not flattering.
I logged every outbound call from my installed skills for 30 days, and tracked which ones I actually used.
- 70% of installed skills made zero calls in 30 days but still held full outbound permission
- 4 skills sent fields from my prompt I did not realize were leaving the machine
- 3 overlapped almost completely. I had been stacking, not replacing
- The only built-in "recommendation" I see anywhere is "top installed", which says nothing about my workflow
The frame I keep coming back to: privacy and recommendations are the same problem on two different clocks. Privacy answers "what is this skill doing right now." Recommendations answer "should this skill still be here next week." Neither is well-served by anything I currently have running.
Real questions for this sub:
- If you have more than ~10 skills installed, what is your idle ratio? Want to know if 70% is the norm or if I am an outlier
- What signal would actually convince you to uninstall a skill you have been holding onto? Last-used date alone feels too crude
- For privacy specifically, would you trust a redaction layer that runs locally before each outbound call, or is "read the source" still your default?
Disclosure since this community deserves directness: I have been working on an open-source tool around exactly these two layers, and would rather have it torn apart here than by strangers. https://github.com/mapick-ai/mapick. Post is downstream of the question though: are these the right two layers to be focusing on, or am I optimizing the wrong thing?