claude skills, explained in plain english.

elisabeth hitz · june 11, 2026 · 6 min read

"skills" is the word anthropic uses, and honestly, it is builder vocabulary. nobody wakes up wanting to buy a skill. what people actually ask me is "can you set up the thing that drafts my replies?" or "how do i make it write like me every time?" a skill is just the answer to those questions. so here is what one actually is, with the jargon removed.

a skill is a written-down way of doing one task

a claude skill is a plain-text file, usually called SKILL.md, that lives in a folder in your claude setup. inside is a description of one task, written the way you would brief a careful new hire: here is the job, here are the steps, here is the tone, here are the rules you never break, here is what good output looks like.

when you ask claude to do that task, it reads the file first and follows it. that is the entire mechanism. nothing is compiled, nothing is hosted, there is no dashboard. you can open the file in any text editor and read every word of the instructions. if the output is wrong, you edit the file once and the fix holds from then on.

this is the difference between prompting and a system. a prompt lives in your head and varies with your mood. a skill lives on disk and does the task the same way on a friday afternoon as on a monday morning. the research on why most AI projects fail points at exactly this gap, and i broke that down in why your ai is not working: the workflow being written down, not the model, decides whether you see a result.

why owned files beat another subscription

most "AI automation" on the market is sold as a hosted platform: your workflow lives on someone else's server, and you pay monthly for the privilege of it continuing to exist. stop paying and the system you depend on disappears with the login.

i hear what people think of that model every day. "we have too many subscriptions going already." "i don't want to be beholden to these subscriptions forever." the median small business already runs about five AI tools, and roughly a third of that spend sits unused within 90 days. the appetite for tool number six is zero.

skill files are the opposite shape:

• you pay once and they are yours. the files sit in your account, on your machine, in your folders.
• they are readable. every instruction is plain english you can inspect before it ever runs. you always know why it did what it did.
• they are portable. the file is a documented description of how your business does a task. even if you switched away from claude someday, that playbook is still yours in plain text.
• they run on the plan you already pay for. the claude subscription you have is the only recurring cost.

interactive vs metered: the part that became urgent this month

that last point got a lot more important. on june 15, 2026, anthropic moves programmatic claude usage, meaning background agents, headless scripts, and automation that runs without you in the loop, onto a separate metered credit billed at API rates. interactive use, the side where you are present and working with claude directly in chat, claude code, or cowork, stays on the flat subscription.

skill files live on the interactive side. you open claude, the skill runs, you review what it drafted. that lane stays flat rate. the always-on hosted automations many agencies sell are the lane that just got a meter attached. i wrote the full breakdown in what the june 15 metering change means for you, but the architecture point stands on its own: a system built from owned files that run interactively is built on the side of the line that stays predictable.

a warning about free skill dumps

because skills are just files, the internet is full of free ones. download with care. snyk's toxicskills research (february 2026) scanned 3,984 publicly shared agent skills from free directories and found 36.82 percent had at least one security flaw, 13.4 percent contained at least one critical issue, and 76 were confirmed malicious, doing things like credential theft. worse, automated scanners caught only a small fraction of the malicious ones, because the attack lives in the instructions, not the code.

the practical rule: only run skill files you can read and understand, from a named human you can actually reply to. every file i sell is written by me and readable line by line before it runs. that is also the standard you should hold anyone to if you hire setup help.

what a working skill stack looks like

one skill is a party trick. a stack is a system. mine, the one i use on camera, runs in this order:

1. the brain file. who you are, your offer, your voice, your rules. every other skill reads it, so nothing sounds like a robot. you can build this one free with the brain builder.
2. the content system. one idea in, a week of posts out, in your voice.
3. the pitch and reply system. inquiries answered fast, in your words, with you approving every send.

each one is the same pattern: a task you do every week, written down once, running forever after. for a business owner the same pattern covers invoice chasing and lead follow-up, which is the version i cover in how to set up claude for a small business.