A skill is the process you keep explaining
Every business has a job you explain the same way every time. Mine was shipping app updates to Apple.
Same checklist. Same traps. I kept re-teaching it to a fresh chat, and the fresh chat kept forgetting.
A skill is where that explanation lives now. It’s a file called SKILL.md inside its own folder. The top holds a short description of when to use it.
Below that are the steps, written the way you’d brief a careful new hire. Not ready to write one? Start with the free AI business starter.
What’s inside one of mine
My ship-ios-app skill exists because Endii, my endometriosis app, went through App Review more than once before it was approved.
Everything that got us through is in one folder:
SKILL.md: the playbook, 133 lines.
references/: five files covering the upload, the review flow, compliance, QA and API hardening.
apps/: one settings file each for Endii, Pause, Timeline and Kiki.
The playbook is the same for all four apps. Only the details change, so those live in their own files.
When Claude ships Pause, it reads the Pause file. It doesn’t guess the IDs from memory.
The skill keeps the scars too. In August, Apple filed one of our builds under an old version number and rejected it.
We never found the root cause. The fix went into the skill as a rule: check the exact file, then upload that same file, never the shortcut.
The next release starts from that lesson instead of from zero.
How does Claude know when to use a skill?
It reads the descriptions first. At the start of a session, Claude sees only each skill’s name and a line or two about when to use it.
The full instructions load when your request matches. That’s why you can install dozens of skills without filling up the conversation.
So the description carries the weight. Mine lists the phrases I actually say, like “ship a build” and “swap the build”.
A vague description means a skill that never gets picked.
Do you need to code to make a Claude skill?
No. A skill is mostly writing. Some include scripts Claude can run, but the core is plain instructions. My idea skill is 49 lines of prose.
It tells Claude to write my idea down word for word before judging it, then push back honestly, then give one verdict: build, park or kill.
The easiest way to start is to do the job with Claude once, then ask it to turn what you just did into a skill.
Read what it writes. Cut anything that isn’t how you actually work.
Skills go stale
A confession. My wrapup skill, the one that saves my progress at the end of a session, still tells Claude to send my next actions to Sunsama.
I cancelled Sunsama in August. The skill only survives because it also says to skip that step when Sunsama isn’t connected.
A skill is a document, and documents age. When your process changes, change the file the same day.
When skills matter, and when they don’t
They matter when you repeat a job and the steps barely change: a release checklist, a weekly review, a brand check before anything goes out.
They matter more on a team, because the process stops living in one person’s head.
They don’t help with one-off questions. And a skill can’t reach a tool Claude isn’t connected to.
It can say “check the calendar”, but with no calendar connection that line does nothing.
Connections are a separate thing, called MCP. Here’s how skills and MCP differ, and what MCP is.
If your team keeps re-explaining the same process, start with training for teams: one shared task, written down so the next person can repeat it.
Official definition: Claude Help Center: What are skills?