The private auto-updating marketplace is the detail I would steal. Shared skills rot silently otherwise: one person fixes theirs, everyone else keeps running the broken copy. I maintain a public creator skill pack solo, and versioning is the problem I have solved worst. Curious how team-2x handles a skill that improves for one team and regresses for another.
Good question! We definitely have a lot of similar skills customised for each team. On the face of it this should be a bad thing - re-use is good, but I'm fine if it means we have small similar skills suited for each team that do the job really well! Not everything needs to be a monolith.
The bad-skills failure mode I keep hitting is a skill that’s too broad, it fires for the wrong task and quietly does something plausible instead of nothing. Curious what your bar is for splitting one skill into two versus just tightening the description.
Mostly it's if the skill "smells" too complex or we see that it doesn't get to the right outcome quickly. Then it's a good time to break things up. We use a mix of user vigilance and transcript mining to do this.
The private auto-updating marketplace is the detail I would steal. Shared skills rot silently otherwise: one person fixes theirs, everyone else keeps running the broken copy. I maintain a public creator skill pack solo, and versioning is the problem I have solved worst. Curious how team-2x handles a skill that improves for one team and regresses for another.
Good question! We definitely have a lot of similar skills customised for each team. On the face of it this should be a bad thing - re-use is good, but I'm fine if it means we have small similar skills suited for each team that do the job really well! Not everything needs to be a monolith.
The bad-skills failure mode I keep hitting is a skill that’s too broad, it fires for the wrong task and quietly does something plausible instead of nothing. Curious what your bar is for splitting one skill into two versus just tightening the description.
Mostly it's if the skill "smells" too complex or we see that it doesn't get to the right outcome quickly. Then it's a good time to break things up. We use a mix of user vigilance and transcript mining to do this.