Claude Skills: 3 real recipes + the rule that makes them fire

Most people use Claude Skills like a ‘remember this for me’ feature. But used right, they become something else entirely. Security boundaries. Guardrails. A logbook of every landmine. Today: three Skills we actually run — and the one rule they all share. The structure is simple. A Claude Skill is just a markdown file. Name and description at the top, body underneath. That’s it. Claude usually only reads the description. When something in the conversation matches that description — then, and only then — it opens the body. That’s the entry level. Today, what really matters: what you put in the body. We’ll see it through three real Skills. Skill number one: cloudflare. Handing Claude a Cloudflare API token sounds harmless. It isn’t. A full-permission token can touch your DNS, your SSL, your Workers, your zone settings — everything. One slip and your site is gone. So our cloudflare Skill has three rules. One: never print the token value in logs or output. Two: any destructive operation — DNS delete, Worker delete — must be confirmed with the user first. Three: the token itself uses a minimal scope. Ours is restricted to DNS edits on indieailab dot com. No Workers edit. No Account settings. Plus a landmine we found in the field: Cloudflare scoped tokens fail the tokens-slash-verify endpoint by design. That’s not a broken token — don’t recreate it. That’s Skill purpose number one — encoding a security boundary. Skill number two: drive. Letting Claude touch your Google Drive is convenient, but risky. Say ‘clean up old files’ and look away, and important documents could be permanently deleted. So the drive Skill spells out the forbidden moves. Permanent delete: never. Move to trash, fine. Overwriting existing spreadsheet data: never. Appending, fine. Restructuring document headings: only after explicit confirmation. There’s a positive procedure too. Read the current state before any write. This alone kills ‘wait, that’s not what I expected’ moments. And a small detail: spreadsheets with multiple sheets must always be told which sheet to write to. No more accidents on Sheet 1. That’s Skill purpose number two — guardrails so the AI never quietly breaks things. Skill number three: kdp. This one’s my favorite. Suppose you publish PDFs to Amazon’s Kindle Direct Publishing. Anyone who’s done it has hit this exact wall: the italic font Helvetica-Oblique simply doesn’t render in KDP’s preview. It’s not in the official docs. You hit it, you search, you wonder, you eventually realize the answer is — don’t use that font. Normally, six months later, when you make another book, you hit the exact same wall again. But if you write it into the kdp Skill, Claude will never pick that font again. Same with: don’t paint background color all the way to the page edge — that’s a margin violation. Things the official docs don’t say. Pain you’ve felt once, never feel again. That’s Skill purpose number three — banked field knowledge. Cry once. Now, the rule I promised. All three Skills have totally different purposes — security, guardrails, field knowledge. But for any of them to actually fire, one thing has to be right. Not the body. The description. cloudflare’s description: Cloudflare API, DNS, SSL, Workers operations. The verb and the object are clear. drive: reading and writing Google Drive, Sheets, Docs. Unambiguous. kdp: producing or uploading KDP books. No drift. If that description had just said ‘code reviewer’, no matter how brilliant the body, Claude would never call it. The description isn’t decoration. It’s the invitation. Skills aren’t a feature. They’re a memory. A security boundary. A set of guardrails. A vault for everything you broke once. Write it once, never repeat the same mistake. One rule: the description is everything.