Claude shipped a feature called Skills, and most people are using it wrong. There’s one specific thing about how Skills work that almost nobody explains correctly — and once you understand it, the whole feature stops being confusing and starts being genuinely transformational. Here’s the concept, why it matters, and the small fix that makes every skill you build actually fire.

The Problem Everyone Hits

Here’s how it usually goes. You build a custom skill — say, one for cleaning up meeting transcripts — and Claude is supposed to use it. You paste in a fresh transcript, and… nothing. No indicator, no tool call, just Claude answering normally as if the skill didn’t exist. The description is clear, you told Claude exactly what to do, and it’s still being ignored. That’s the moment most people decide Skills are broken. They’re not — the skill is just built on a wrong assumption that almost every tutorial repeats.

The One Thing Nobody Explains: The Description Is the Trigger

A skill file has two fields at the top — a name and a description — followed by the instructions and the playbook. The instinct is to make the description as detailed as possible, explaining what the skill does and the how and the why. That’s the mistake. The description field isn’t a description. It’s the trigger. Read the example wording closely — “use when asked for updates or progress summaries” — and it clicks: that field exists to match the request, not to explain the skill.

How to Fix a Skill That Won’t Fire

Every time you start a new conversation, every skill’s description loads into Claude’s context and gets scanned against each message, looking for a match. When something you say matches, Claude pulls in that skill’s full instructions and playbook. If nothing matches, the skill stays asleep. So the description has exactly one job: match the request. A long, careful, paragraph-style description is actually worse than a short one, because it’s full of words that don’t match. Rewrite your descriptions as trigger phrases — one sentence containing the words or situations you’d actually use in the moment, not the words you’d use to explain the skill to a coworker.

What This Tells Us About Where Claude Is Going

This is why I don’t think Skills are a custom-GPT clone. A custom GPT is a destination — you go to it, use it, and leave. Skills are routes. You stay where you are, and Claude figures out which one to walk down based on what you’re doing. The tool switches itself mid-conversation without you noticing. The description isn’t telling Claude what the skill does; it’s telling the router when to wake the skill up. Reframe it that way and the whole feature makes sense.

There’s one piece I left out on purpose, because it deserves its own video: skills can include code that actually runs — real scripts Claude executes when the skill fires. That’s where this stops feeling like a prompt and starts feeling like an operating system. More on that soon. If this was helpful, leave a thumbs up and check out the related Claude video above.