I have a confession. For about two years my best prompts lived in a notes app called "PROMPTS FINAL v3."
I'd scroll, copy, paste, then burn five minutes fixing whatever didn't fit. Half the time I grabbed last month's version instead of the good one. My team kept their own notes apps, and of course nobody's matched mine.
I'd bet your shop runs AI about the same way. A dozen laptops, a dozen private stashes.
Last week the plumbing for fixing that got a lot more serious. On October 22 at AGNTCon + MCPCon North America in San Jose, the Linux Foundation's Agentic AI Foundation announced four new projects, and the one I care about for you is Agent Skills, donated by Anthropic. OpenAI chipped in Open Responses, Hugging Face brought OpenEnv and Alibaba brought OpenSandbox. David Soria Parra of Anthropic said they're donating Skills "to ensure that anyone building with skills can shape what's next."
Meanwhile Google is killing off Gems and replacing them with Skills. So the two biggest names in your toolbox now point the same way.
Your best prompts are about to get a real home. I'll walk you through building one.
What a skill actually is
You have to go find a prompt and paste it. A skill gets picked up by the AI on its own when the job calls for it.
Under the hood it's a small folder. Inside sits a file of instructions, usually a few examples of good work, maybe some reference files (your price sheet, your brand voice notes, your refund policy) and now and then a script that handles a specific calculation or formatting job. The Agent Skills format describes it as a way to bundle instructions, scripts and resources that agents load on demand, so specialized know how travels across products and platforms.
People are using it, too. According to AI Weekly, there are about 1 million skills published and 280 million installs since December 2025. Somebody's science fair project doesn't get those numbers.
Four reasons it beats a saved prompt.
Consistency. The skill carries the example and the quality bar every single time, so the good version is the only version anyone sees.
Team sharing. One person writes it and everyone uses that same copy. Your newest hire gets your best thinking in week one.
It loads itself. The AI reads a short description of every skill it has and decides which ones fit the request. You don't have to remember it exists.
It travels. The format is turning into a shared standard, so a skill you write today won't be stuck inside one vendor's walled garden.
Pick your first three
Don't start by turning your whole business into skills. Start with the jobs you've explained out loud more than three times.
Every company has tribal knowledge. It lives in one person's head and comes out as "no, we always do it this way." A few that show up almost everywhere:
The weekly numbers recap. Somebody pulls leads, jobs booked, revenue and cash, then writes the same three paragraphs for the owner every Monday.
The proposal in your voice. The owner writes great proposals. Everybody else's read like a rental car agreement.
Review replies. Five stars, two stars, the guy who's mad about parking. Each one needs a reply that sounds human and doesn't land you in legal trouble.
Job posts. You hire twice a year and reinvent the post from scratch every time.
SOP writing. Turning "watch me do it" into a written procedure someone else can follow.
A quick example, and it's a made up one. Picture a six person home services company: owner, office manager, two techs, a dispatcher and a part time bookkeeper. The office manager, call her Dana, is the only one who knows how the owner likes review replies written. Only the bookkeeper knows which numbers the owner actually reads on Monday. And nobody but the owner writes estimates that close.
If Dana takes a week off, reviews sit. If the bookkeeper quits, the Monday recap dies.
So they build three skills: review replies, the Monday recap and estimate follow ups. What lived in three people's heads now sits in folders anybody can open. That's most of the work, honestly.
Anatomy of a skill that works
Most bad skills fail in the same spot, and it's usually the description, of all things.
A name. Short, lowercase, one word or close to it. You'll see it in lists.
A one line description that says WHEN to use it. This is the part the AI reads to decide whether to load the skill. Write it like a trigger, not a slogan. "Use when replying to a customer review on Google, Yelp or Facebook" works. "Helps with customer communication" is useless, because everything is customer communication.
The steps. Numbered, in order, written for a sharp new employee who's never met you.
The quality bar. What does done look like? Length, tone, what has to be in there.
An example of good output. I'd take one real example over a paragraph of adjectives every time. If you only add one thing to your prompts this year, add this.
What NOT to do. Every business has landmines. Never offer refunds in public, say, or never promise a callback time.
Reference files. The stuff the AI should look at but you don't want crammed into the instructions. Price ranges, service area, warranty terms.
Keep the main instructions tight. A 40 page handbook gets skimmed, by the AI and by people.
Let the AI interview you
Writing a skill from scratch is a chore. So don't. Have the AI interview you and draft it from a job you already did well.
Grab one past example where the output was exactly right. A proposal that closed, or a Monday recap the owner actually liked. If the knowledge lives in a meeting, record the next one in Fathom and use the transcript as raw material. A lot of tribal knowledge only comes out of somebody's mouth when they're explaining it to a new hire.
I want to turn a job I do over and over into a reusable AI skill. Interview me first, then draft the skill. The job: [describe it in one sentence] Here's a past example of this job done exactly right: [paste the example] Ask me questions one at a time, no more than 8 total. Cover: when this job comes up, who it's for, the steps I follow, what good looks like, mistakes I've seen people make, anything that should never happen, and what reference info the job needs. When you're done asking, write the skill with these parts: a short lowercase name, a one sentence description that starts with "Use when" and names the exact situations that should trigger it, numbered steps, a quality bar, a "never do this" list, my example as the example of good output, and a list of reference files I should attach. Keep the main instructions under 400 words. Flag anything in my example that looks like a one time exception rather than the rule. |
Don't skip that last line. Your great example probably has a quirk in it, like a discount you gave one customer. You don't want the skill learning that as policy.
FROM THE AI NEWSROOM The Skill Starter Kit Ten ready to use skills for the jobs every small business repeats, the anatomy template, the skill builder interview prompt, a five case test log and a simple system for keeping them current. Free for AI Newsroom readers. Reply SKILLS to this email and I'll send it straight over. Reply SKILLS to get the kit |
A skill you can steal
Below is a complete review reply skill, laid out the way a SKILL.md file is. The first two lines are the frontmatter: the name and the description. In the actual file those two lines sit between a line of three dashes above and a line of three dashes below. I left the dashes out so your email client doesn't mangle them. Swap in your business details and you're done.
name: reviewreply description: Use when writing a public reply to a customer review on Google, Yelp, Facebook or any review site, for any star rating. Also use when the user pastes a review and asks how to respond. # Review replies for [Business Name] You write public replies to customer reviews for [Business Name], a [type of business] serving [service area]. Every reply is read by future customers more than by the reviewer, so write for them. ## Steps 1. Read the review and identify the star rating, the specific thing they praised or complained about, and any names, dates or job details. 2. Thank them in the first sentence, using their first name if it's shown. 3. Mention one specific detail from their review so it's obvious a human read it. 4. For 4 and 5 stars: keep it warm and short, and invite them back for the next job. 5. For 1 to 3 stars: acknowledge the problem without arguing, say what we're doing about it, and move the conversation offline with this line: "Please call [Name] directly at [phone] so we can make this right." 6. Sign off as "[Owner first name], owner." ## Quality bar 40 to 90 words. Sounds like a friendly owner, not a corporate apology. One specific detail from the review. No more than one exclamation point. ## Never do this Never admit legal fault or mention insurance. Never offer a refund, discount or free service in public. Never name an employee in a negative review. Never argue facts, even if the customer is wrong. Never copy the same reply twice. ## Example of a good reply (2 stars, late arrival) "Thanks for the honest feedback, Mike. You're right that we showed up two hours past the window, and that's not the standard we hold ourselves to. We've changed how we schedule afternoon calls so it doesn't happen again. Please call Dana directly at [phone] so we can make this right. Tom, owner." ## Reference Use the file service_area.md for towns we serve and service_list.md for what we do and don't offer. |
Notice how boring the description is. I did that on purpose. Boring and specific gets triggered at the right time. Clever gets ignored.
Test it on five old jobs
A skill you haven't tested is just a prompt with better posture.
Pull five real past cases. Five reviews you already answered, five past weeks of numbers, five old estimates. Run the skill on each one and put its output next to what you actually sent.
Grade each output: send as is, light edits, or trash. If three of the five need heavy editing, the skill isn't done yet. Most of the time the fix is a sharper step or a better example. Piling on more words rarely helps.
Then test the trigger, which is where most people get burned. Ask for things that should load the skill and a few things that shouldn't.
It doesn't trigger when it should. Your description is too narrow or too vague. Add the actual words people use. If your team says "respond to this Yelp thing," put Yelp in the description.
It triggers when it shouldn't. Your description is too broad. Say your review skill fires every time someone asks for a customer email. Add "public review" and "not for private emails" to the description and try again.
Budget about an hour per skill for the first round, then keep tightening it as real jobs come through. A new hire gets a few weeks of feedback too.
Where your skills live
Claude. Claude supports skills directly, so it's the most natural home for the format. Upload the folder, and Claude decides when to use it based on the description.
Gemini. Google is replacing Gems with Skills, which it describes as reusable custom instructions. You call one by typing a forward slash, you can combine several in one chat (Gems couldn't do that), and Gemini can spot when a skill would help and use it without being asked. Reference files are supported, with sharing and attaching Drive files coming in the following weeks.
If you've got Gems, pay attention to the timeline. Since October 13 you can't create or edit Gems. Personal account Gems migrate to Skills automatically on November 17. Workspace business and enterprise Gems get removed in March 2027. Just because a Gem migrates on its own doesn't mean it'll come out the other side any good. Open each one before then, decide if it stays, and rewrite the description so it says when to use it. Most Gems were written as personas, and a persona makes a lousy trigger.
Everything else. For tools that don't support the format yet, paste the instructions into a project or the custom instructions area and attach the reference files there. You lose the automatic loading but keep the consistency. Notion says its MCP overhaul will add skills via MCP soon, so I'd expect this list to shrink.
Keep a master copy of every skill in one shared folder, outside any single tool. Tools come and go (ask anyone who built a pile of Gems), and that folder is the thing you actually own.
Read it before you install it
Skills are spreading fast, and people are sharing them. This week's SuperCMO AI launch on Product Hunt packages marketing expertise as open source Claude skills. Good sign for the format, if you ask me.
But remember what's inside a skill folder: instructions, files and sometimes scripts. A script runs code. A sloppy or malicious skill can tell the AI to do things you'd never approve if a stranger asked you directly.
So read a skill the way you'd read a contract from a vendor you've never heard of. Open every file before you install it. If there's a script and you can't tell what it does, ask the AI to explain it line by line, or skip it. Stick to sources you trust, and when in doubt, use someone else's skill as inspiration and write your own.
FROM THE AI NEWSROOM The AI Business Accelerator For operators who would rather build it with someone than build it alone. The full stack, installed alongside you, with the decisions made in the room instead of guessed at. Ninety seven dollars. Join the Accelerator for $97 |
This week
Write down every job you've explained to someone more than three times this year. Circle three.
Find one past example of each job done exactly right.
Run the interview prompt on your first one and get a draft skill.
Test it on five real past cases and fix the description until it triggers when it should.
If you've got Gems, open every one before November 17 and decide what stays.
Put the finished skills in one shared folder and tell your team they exist.
Jordan
The AI Newsroom | Practical AI for people with a business to run.