You've Heard of AI "Skills." Here's How to Actually Use Them.

Happy Wednesday! This week:

Let’s get started!

A few weeks ago I got on a call with the president of a comms association chapter who wanted to build an AI tool for her clients. She’d assumed she needed to hire a developer, and had been hunting around for a bootcamp to avoid it. What she actually needed was a text file — but didn’t know it, because all she had heard about were “Skills.” And that sounded scary.

So let’s demystify this, because “Skills” is about to become one of those words everybody nods along to and nobody asks about.

A Skill is a written-down procedure you hand to the AI. That’s really it — a text file with a bunch of instructions for your LLM. Claude can read them automatically — you’ll need to point other LLMs to the files if you want to use, say, ChatGPT or Perplexity instead, but we’ll get to that later in this email. For our purposes here, let’s assume you’re using Claude.

Skills feel kind of magical because you don't paste them into a chat and you don't remind Claude that they exist — when you ask for something that matches what the file describes, it gets picked up and followed. That's the difference between a Skill and a really good prompt. A prompt lives inside one conversation and dies with it. A Skill sits outside the conversation, and every conversation can reach for it, which is why the tenth time you run the job comes out like the first.

This is actually super-easy because you don’t have to learn how to do it — your LLM does it for you. You just need to realize that it’s an option, and a good one.

The next time you finish a piece of work with the AI — a coverage scan, a media list, a round of edits — just say turn what we did into a Skill. It has the whole session in front of it, so it already knows the steps, the source material, and every correction you made along the way. You'll have a draft in about thirty seconds. You’ll need to save it, though — which is easy. It’s a one-time setup — go to Settings → Capabilities and turn on "Code execution and file creation." Skills don't run without it, and this is the most common reason someone tries this and nothing happens. (On a Team or Enterprise plan your admin flips that one. And if you can’t find the setting, just ask Claude for help.) Once that's on, Claude saves the file straight to your library — you'll find it under Settings → Capabilities → Skills, where you can switch each one on or off.

Because it’s just a text file, you can manually adjust it if you want — read it, cut what's wrong, and add what it couldn't have known — the rules you never said out loud. But you don’t have to manually adjust it this way — I rarely do. Because there’s a better way:

Ask Claude to check it for you. I really like this trick. Open a brand-new chat conversation and ask Claude to read the Skill and tell you what it does and when it would use it. (And yes, the new session is important — the session that wrote the Skill remembers everything you did together, so it'll quietly fill in gaps from memory and the file will look better than it is.) Read its response, tell it to change what’s wrong, and add whatever needs to be added.

I’ll use a real one as an example. I built myself a media intelligence system — a reporter database I can research into and pitch out of — and the piece that finds new journalists is a file called Reporter Discovery. (Creative name, I know.) It’s 64 lines long and has three parts — a name, a description, and a step-by-step guide for how the LLM is going to use it.

The name. It’s the name. No big deal, and if you give Claude access to all of your Skills, you don’t even have to remember the actual name because it’s smart enough to know what you’re talking about (usually). Sometimes I say I want to use my “media research” tool and Claude knows I’m talking about the Reporter Discovery one. If you’re using a Claude Skill with ChatGPT or another LLM, the name matters because you’ll have to manually point the LLM to that file. So use something meaningful to you.

The description. This spells out when Claude should use the Skill — and I think it’s the part that matters most, because it’s how the assistant knows what to do instead of improvising. Mine says to run it when I want journalists found for a given beat or client, for example. It doesn’t have to be complicated. But it’s a good practice to ask Claude or your LLM of choice to tell you what the Skill is supposed to do — you don’t want it guessing.

The steps — how the job gets done. This is the bulk of it, and it’s just how you’d explain the work to a competent new hire. Which searches to run, and in what order. To check the existing database first so we don’t create duplicates. To confirm a reporter is actually still at that outlet before writing them down. Whatever guardrails are important. I like the “AI as a really capable intern” analogy — the more instructions and safeguards you put into place, the better results you’ll get. This is even more important when Skills are involved.

And that’s all the Skill is — instructions. The Skill is the recipe, not the pantry. A recipe card doesn’t reprint your spice rack — it tells you to go get the onion powder. The Skill doesn't hold the information Claude needs, either — it just tells Claude where to go get that onion powder.

Most people are doing one of two things.

You’re re-typing your instructions every time. You’ve got a prompt that works, it lives in a Google Doc or in your head, and every week you paste it in again with slight variations. It works fine, but that’s a lot of copy/pasting and relies on your memory. Even worse, it’ll also never improve, because there’s no single copy to improve.

Or you’ve built a project and filled it with documents. This is where I see the most frustration. A project is a container — you load it with brand guidelines, the style guide, past coverage reports, and expect good output to fall out. But a pile of reference material tells the model what your work looks like. It doesn’t tell it how the work gets done. Those are different problems, and comms work is often centered on the second one.

Skills are a Claude thing. Your file isn’t — and that’s good if you need to use another LLM.

When using Claude, it automatically notices when your request matches a Skill, finds the right file, and loads it without you doing anything. ChatGPT and Perplexity don’t have that piece — so there are some limitations. But the file is plain text, and the file is the actual asset. Case in point: my media system doesn’t use the Skills feature at all. Those procedures are just files sitting in a folder, and I open the one I want and say run it. It works exactly as well. The automatic retrieval is a convenience — and a very nice one at that — but not a requirement.

Which means the same file goes wherever you do. Drop it into a Custom GPT’s instructions and you’ve got it. Paste it at the top of a Perplexity thread and you’ve got it with one extra step. And if your company switches vendors — at some point it will — you carry your process across in an afternoon. A project doesn’t come with you. Three months of refinement stays stuck in a container you don’t control. And yes, that also means you can share Skills with your colleagues, but make sure you don’t have anything private in the files first, of course!

For everyone reading this where IT has said no to most of it: you don’t need the feature at all.

A Skill is just a file of context, and if you can’t save one, you can paste one. Keep it in a doc, paste it at the top of the chat, and you’re there. That’s the entire workaround.

Your next steps: Next time you do something repeatable, ask your LLM to create a Skill (and then vet the output per my instructions above). I’ll make it super-simple for you — here are the prompts you can use to generate the Skill or ChatGPT/Perplexity equivalent:

For Claude:

The portable, LLM-agnostic version:

That’s it. Give it a try and let me know how it goes.

Posting everywhere means rewriting one idea six times, so you post to one, or none. SureThing turns one idea into native posts for every platform.

Put It on Autopilot

Last week the billionaire financier Stanley Druckenmiller published a Wall Street Journal op-ed arguing the Treasury Department should stop trying to push down long-term bond yields. The argument isn’t what got people talking. The prose was — lines like “This wasn’t liquidity management, it was price management” set off everyone’s detector. NOTUS reporter Jeff Stein called him up to ask, and Druckenmiller said yes, he’d used AI.

Normally that’s the moment we get an apology, an editor’s note, or a quiet takedown. Nope. Instead, as Joshua Benton reports at Nieman Lab, nobody blinked. “I write everything using AI now for the same reason I use a calculator when I do math problems,” Druckenmiller said. Journal editorial page editor Paul Gigot backed him fully, and in a follow-up column said no disclosure was necessary — “it wouldn’t be honest for me to grandstand that our contributors never use a tool like AI. Mr. Druckenmiller’s sin against the media pharisees was admitting its use.”

And then Gigot reached for his best defense, which happens to be us. Plenty of op-eds, he pointed out, are actually written by PR staffers rather than the person in the byline — the same way we all accept that politicians have speechwriters. So how is AI any different?

Benton went and asked. He contacted opinion editors at 24 prominent U.S. newspapers, and not one of the editors who responded was willing to follow the Journal. The Los Angeles Times, Chicago Tribune, Denver Post, Atlanta Journal-Constitution, USA Today, Boston Globe, Philadelphia Inquirer, Seattle Times and New York Times all ban AI-generated submissions outright, most of them in writing on their submission pages. The San Francisco Chronicle’s Matthew Fleischer put it simply: “Our audience is paying for the words and thoughts of human beings.” The Tribune’s Chris Jones was blunter: “We are looking for human opinions, crafted and honed by humans. Period.”

But most of us are using AI to some degree — maybe not to write, but at least to catch typos, right? So we have a real gap: our profession is quietly using these tools to assist in placed thought leadership, and the outlets we place it in are hardening against exactly that — several now asking contributors to attest they didn’t.

I think the answer is simple, though. Gigot’s ghostwriting analogy is wrong — it gets the value backwards. Ghostwriting was never a loophole — it’s a value-add. When a comms person drafts an executive’s op-ed, they’ve sat with that executive, argued with them about the third paragraph, and can defend every claim in it — because a human being with a job on the line signed off on all of it. AI skips that entirely. What editors are policing isn’t who typed the words. It’s whether anyone can still vouch for them.

And I think the Druckenmiller imbroglio accidentally proves the point. The wonky, robotic AI voice didn’t do Druckenmiller any favors — if your audience is talking about who (or what) wrote your thought leadership instead of the thinking itself, you’ve failed. Leaving that kind of language in a piece is a big mistake; as communications professionals, it’s our job to ensure that it doesn’t happen.

If you’ve ever asked ChatGPT who covers your beat and gotten back a list of reporters who left the outlet two years ago, you already know the problem. LLMs are often bad at curating recent news — outdated, spotty on coverage, and pointing at sources you can’t verify.

Perigon makes a tool that attacks this problem, and next week, they’re hosting a free webinar. Joshua Dziabiak, Perigon’s founder and CEO, is walking through how to connect their news API and MCP server to the AI tools you already use — Claude, ChatGPT, or an agent you’ve built yourself. The part worth showing up for is what you can pull once it’s wired in: articles, journalists, sources, companies, sentiment, etc. — it turns a pile of tedious work into questions you just ask.

It’s 45 minutes on Tuesday, Sept. 8 at 2:00 PM ET, co-hosted by my friend Stephen Karaolis. Everyone who attends live gets extended Perigon API access afterward so you have room to actually try what you learned, too, so there’s practical value here.

Save your seat →

Whether you're trying to build systematic AI adoption across your team or focused on AI search visibility for your brand or organization, I’m here to help. Reply to this email — let’s talk.

Most AI resources are scattered, untested, or buried behind a paywall.

The Shift puts it all in one free newsletter, covering what’s new, and what practically works under 5 minutes a day.

A vault of 3,000+ vetted tools so you’re never guessing. A 1000+ prompt library for every use case you’ll actually hit and free AI courses that make sure nothing stays theoretical.

Plus, right now, 3 subscribers win a free 1-year Claude Pro subscription just for signing up.

Subscribe to The Shift

Looking for a role at the intersection of communications and AI? Here are some opportunities to check out:

Was this forwarded to you by a friend or colleague? Want to get this in your inbox next week? Click here to subscribe!

Quick question: how can I help?

What workflows are you struggling with? Where does AI still feel mysterious or overwhelming? What has worked that you’d like to share with others?

I’m a builder, and I’d love to help you and the rest of The Comms Stack community find great new ways to use AI.

Reply and tell me.
I read every response.

Even a one-sentence reply helps. For example:
“I wish AI could help me with ______.”

Until next Wednesday,

Dan

Advertise | Subscribe