Ask HN: Are most agent skills just fancy system prompts with a name?
2 comments
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.
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.
One issue I’ve run into is that different models (e.g., Codex vs. Claude/Opus) often require slightly different configurations. This leads to duplicated files like .codex/skill.md and .claude/skill.md, which adds maintenance overhead.
For example, when working in non-English contexts, I need to ensure the AI doesn’t generate non-English inline comments. Constraints like this often have to be repeated across model-specific configs.
In my case, since enforcing English comments is a priority, I tend to put this in skill.md. Otherwise, I would place it in agent.md.
Because of this, I prefer using a single AGENTS.md as a global configuration layer. It’s simpler to manage and reduces duplication.