I Found Out My AI Skill Was Broken, And It Was "Working" the Whole Time

So I've had this Morning Briefing skill running in Claude Code for a while now. Every morning, it searches the web and gives me a full news briefing — Nepal news, international headlines, crypto, oil prices, the works. It was working, so I never touched it. Turns out working and well-built are not the same thing.

I installed a Skill Builder skill this week — basically a skill that helps you build new skills but i used it to audits your other skills against best practices. Ran it on my morning briefing and it found a bunch of stuff I never would've caught on my own. The description was 12 lines long (Claude has a shared budget for all skill descriptions — mine was hogging it). There was no argument hint so the slash menu showed nothing useful. And the big one — disable-model-invocation wasn't set, which means Claude could randomly fire off 10 web searches in the middle of a conversation without me asking. Not great.

Fixed everything in like 10 minutes. The briefing output looks the same — because the content was already solid — but the skill is now lean, clean, and won't go rogue on me and sake tons of token.

Honestly the lesson here is simple: your Claude skills are not just prompts, they're more like mini software. And software needs maintenance. If you've built skills and never audited them, Skill Builder is worth installing. Drop a comment if you want to know how to set it up.

11 comments

I Found Out My AI Skill Was Broken, And It Was "Working" the Whole Time