Free AI skills are everywhere right now.
Comment “SKILL” on a LinkedIn post and someone will send you a folder that promises to write your proposals, run your outbound or summarize your inbox. Upload it, type a slash command, done.
Here’s the harsh reality: a skill isn’t a template. It’s a set of instructions your AI follows – with your accounts, your files and your usage plan. Installing one is closer to handing a stranger a key to the office than downloading a font.
That doesn’t mean don’t use them. We downloaded one ourselves. It means read it first, rebuild what you keep, and prove it works before you trust it.
First, what a skill actually is
A skill is a folder of instructions (and sometimes small programs) that teaches an AI tool like Claude to do one job the same way every time. The AI reads a one-line description to decide when to use it, then follows the instructions inside.
Two things follow from that. The instructions run with whatever access your AI has. And the description decides when it runs – so a badly worded one can fire when you didn’t ask for it.
Before you download: six things to check
- Who wrote it, and who maintains it: a skill nobody maintains breaks quietly the next time the AI model or a connected app changes. If you can’t name the author, assume nobody is fixing it.
- The license: this one bit us. The council skill we started from had no author or license recorded anywhere. That’s fine for poking around. It’s not fine for anything you’ll sell, install for a client or build your business on. Some licenses (MIT, for example) allow commercial use and changes. Many don’t, and “no license” means you have no stated right to reuse it at all.
- Every file, not just the summary: skills can include scripts that actually run. Anthropic, the company behind Claude, recommends installing skills only from sources you trust and auditing anything else before use. If a folder contains code you can’t read, have someone who can read it look before you install it.
- What it can reach: a skill that “summarizes your inbox” runs with the same email access that can send, forward and delete. Look for explicit rules that it never sends, posts, publishes, deletes or spends money without asking you first. No rules, no install.
- Whether it treats outside content as information, not instructions: a web page, email or document can contain text written to hijack your AI (“ignore your previous instructions and…”). This is called prompt injection. A well-built skill says, in plain words, that nothing it reads is a command.
- Whether you already have one: we once had five overlapping council add-ons installed. Your AI chooses between them from their one-line descriptions, so five similar descriptions is a coin flip. Duplicates don’t add capability. They add confusion.
Modifying a skill: make it yours, properly
We kept the idea behind that council skill and rebuilt nearly everything else. Here’s how to modify a skill without creating a mess you can’t maintain:
- Keep one master copy outside the AI tool: ours lives in our own Google Drive. The copy installed in Claude is always rebuilt from it. If the AI tool disappears tomorrow, you still own the skill.
- Version it: use three numbers. The first goes up when what the skill produces changes, the second when the steps improve, the third for wording fixes. When something breaks, you’ll know exactly which change did it.
- Keep your company facts out of the skill: the skill holds the steps and the judgment. Your names, offers, voice and proof live in your own company documents, and the skill points to them. One skill, many uses – and no “tailored copy” drifting away from the original.
- Write the guardrails in: never send, post, publish, delete, log in or spend without approval. A skill that runs while nobody is watching needs the rules in the instructions, not in your head.
- Write the fallbacks in: what should it do when a connection is missing or an input is wrong? Without a fallback, the AI improvises – and improvising is expensive and unpredictable.
- Keep it lean: every word in a skill is loaded every time it runs. Explanations and history belong in a separate guide for people, not in the instructions the AI pays to read.
Testing: prove it works before you trust it
Most free skills have never been tested by anyone but the author, on the author’s account, with the author’s data. Yours is different.
Write the tests down before you run them:
- Two or three normal cases: the exact request, the setup, and what a pass looks like – written so a non-technical person can check it by reading the output.
- At least one failure case: a missing connection, a missing input, bad data. The skill should fail safely and tell you why, not guess.
- A fresh-account case: run it somewhere with none of your company context. If it suddenly starts mentioning your company or your clients, something is hard-coded that shouldn’t be.
- Measure the cost: our first cost estimate for the council skill was off by about ten times. We only found out because we measured it. A quick run used about 5% of a five-hour usage window on our Claude Team plan; a full run used no more than 13%.
- Re-test after every AI model change: a skill that passed last month can behave differently after an update. We re-test monthly in our own account and after every model change.
- Keep the test log somewhere you can edit: we originally kept ours in a file our Drive couldn’t edit in place. Now every skill’s test results live in one document per skill, with the date, what was tested and PASS or FAIL.
Documentation: if you can’t explain it on two pages, don’t let it run unattended
Every skill we build now ships with three pieces of documentation:
- A guide for people: what it does, who uses it, how to start it, what it needs, what it costs, where it breaks and the legal notes.
- A record: the version, the change history and the test log.
- A one-to-two page descriptor: written for someone who has never opened the skill. Ours covers:
- What it does, in plain words.
- How it starts – by hand, automatically or on a schedule – and the exact words that start it.
- Your part, and how much of your time each run takes.
- What you get back and where it shows up, and what it never does without your approval.
- How it works, in six steps or fewer.
- What it needs, and what happens when something is missing.
- What a run costs, measured, with the date and the plan it was measured on.
- Where it lives, what data it reads and who maintains it.
- Its limitations and legal notes.
Writing the descriptor is a test in itself. Every time we couldn’t explain a section in plain words, we found something in the skill that wasn’t actually decided.
Maintenance: skills drift
A skill isn’t finished when it works. Models change, connected apps change and people edit things.
We run a weekly check. Does every installed skill have a master copy and a record? Does the installed version match the master? Does every skill marked ready have a passing test on file? Anything that fails gets flagged before it causes a problem, not after.
Six questions to ask before anyone installs a skill for you
Whether it’s a freelancer, an agency or your own team:
- Where is the master copy, and do we own it?
- Who owns the skill, and what happens to it if we part ways?
- Can I see the test log, with dates?
- What does a run cost, measured, on a plan like mine?
- What will it never do without asking me?
- What happens when a connection is missing?
If the answer to any of these is a shrug, that’s your answer.
Final thought
A free skill isn’t free. You pay for it in access, in usage and in the cleanup when it does something you didn’t expect.
Read it. Rebuild what you keep. Test it. Document it. Or make sure whoever installs it for you does.