HN user
Honestly, a lot of them are.
We ran SkillCompass across 881 ClawHub skills and abt 46% scored poorly on functional depth, meaning they tell Claude what the skill is but never what to actually do. We kept seeing the same pattern: a name, a vague description, maybe a table that just repeated the skill name in every row.
Claude with the skill vs. Claude without it: behavior was basically the same.
If you say it’s like opening a loot box in csgo, then I totally get it. i used to be pretty good at that game, haha
Project name: SkillCompass
Project description: A tool for testing, diagnosing, and improving AI agent skills. It helps make skill quality easier to inspect, compare, and improve over time instead of relying on guesswork.
Would also love people to just try it and tell us what feels useful vs. rough. Early feedback from people actually working with skills is probably the most valuable thing for us right now.
My current split: Claude for code, Gemini for harder reasoning, ChatGPT for more structured output.
ChatGPT is still useful, but mostly for tasks where formatting, organization, and response shape matter. If I’m judging mostly on raw capability, probably rank Gemini above it.
Sometimes it feels like vibe coding lowers the cost of creating new skills so much that we end up with skills for making skills, and then more skills for evaluating those skills.
At some point the question stops being “how do we evaluate all this?” and becomes “did all of this need to exist in the first place?