For Episode 7 of One Shot Show, Ilia Karelin joined Dheeraj Sharma (GenAI Unplugged) and me as our AI expert guest to walk through Claude Skills live.
Ilia writes Prosper, the newsletter about AI and software that gives you an unfair knowledge advantage. His post on the 3-document system for AI memory is still one of our most-shared pieces, followed by “When NOT to Use AI,” and, most recently, “Claude in Chrome,” “Grok 4.2 Agents Updates,” and “Anthropic Q1 updates.”
Here are Ilia’s latest three posts that you might want to check out:
He came on to answer the question I keep getting from readers:
“Should I be building Claude Skills yet, and how do I know which ones are worth the effort?”
Before we dive in, let’s understand what Claude Skills are.
💡 Quick context if you’re new to Skills
A skill is a reusable instruction set Claude follows when relevant. You write it once as a folder containing a SKILL.md file and any reference docs or templates. Claude pulls the skill in automatically when your message matches its description, or you can invoke it by name with a slash command. Think of it as an SOP that runs itself.
The reason to care is simple. Most of us burn a meaningful chunk of every Claude session re-explaining the same context: voice rules, audience details, examples, banned phrases. Skills move that setup out of the prompt and into a file Claude reads on its own. You build the skill once, and the work compounds from there. The same skill works in Claude Desktop, Claude Cowork, and Claude Code, so the time you put in once pays off across every interface you use.
It’s a fair question. Anthropic shipped Skills 2.0 quietly a few weeks back. The skill creator got dramatically better. People started posting screenshots of folders with 50+ skills. Meanwhile, most people I talk to still have zero.
I have over 30 skills in my own setup. Every single one came from the same trigger: I caught myself doing the same thing manually for the third time in a week, got tired of retyping the setup, and built a skill so I’d never have to type it again.
That’s the rule. Friction first. A skill earns its place when I’m already tired of doing it manually. You shouldn’t build a skill just because it looks cool. Cool won’t lead you to anything useful if you don’t actually use it or care about it.
So I asked Ilia the question I’d been sitting on: how do you decide what becomes a skill? He showed me the question he asks himself before he builds anything:
“What am I doing too often manually?”
That’s the rule we landed on during the live, and it’s the rule keeping my skill folder useful instead of bloated.
How to spot a skill candidate
Dheeraj put the rule in an even tighter form during the live:
“If I have to repeat something like a long prompt three times a week, that’s a pattern. Then I convert it into a skill.”
Three times in a week. That’s the threshold I use too.
Once I started watching for it, it became obvious whenever I found myself:
Explaining the same writing rules.
Pasting the same platform formatting instructions.
Asking for the same type of critique.
Turning one newsletter into the same set of social formats.
Running the same research process before writing.
Those are signals of skills worth turning into systems.
But spotting the candidate is the easy part. The harder craft is evaluating whether a skill actually produces better output than a prompt would on its own. As I said during the show:
“Your time is spent not building the skill. Your time will be spent evaluating the skill output.”
That’s where the episode shifted from “what is a skill” into “how to build one that holds up.” Ilia built one live, showed how the new skill creator runs evals, then demoed a real research skill he uses for running his newsletter.
What will you get inside of post
The rest of this post walks through that live build for AI Maker Lab readers. You’ll get:
The full session recording. The whole live, including the second-opinion skill built from scratch, the Fast Research demo Ilia ran from his phone, and the audience Q&A on Dispatch and memory.
Ilia’s Fast Research skill, packaged for install. Drop the folder into your Claude setup, give it a topic, and get a sourced research brief saved as a Markdown file in your project. The same skill he ran live on the show.
A build walkthrough you can run tonight. The four-element prompt structure that gets cleaner output from the skill creator, how to read the eval panel before a skill ships, when to trigger skills with plain English versus a slash command, and how to set up Fast Research’s first-run interview so the briefings come back filtered through your editorial lens, not Ilia’s.
If you’ve been thinking about Claude Skills for a while and keep putting it off, the recording plus the skill plus the walkthrough should compress the first month of fumbling into a Saturday afternoon.
What the Skill Creator actually is
A quick note on the tool that did most of the work in this episode, since the rest of the post assumes you know what it is.
The Skill Creator is a meta-skill: a skill whose job is to build other skills. Anthropic ships it pre-installed across Claude Desktop and Claude Cowork so you don’t have to download or configure anything. It’s already there the first time you open Claude. You can see the details inside the Customize > Skill menu.
But if you use Claude Code, you’ll need to install the plugins first. I wrote an in-depth post about it here: From Blank Folder to Working System: How to Set Up Any Project in Claude Code.
To trigger it, mention it by name in your prompt:
“Help me create a [name] skill using the Skill Creator.”
That’s the whole invocation. When you mention “Skill Creator” in a prompt about building a skill, Claude loads the meta-skill and walks you through the build flow: scaffolding the SKILL.md, generating test prompts, running evals against them, and saving the result into the right folder.
Once you know the Skill Creator exists and how to call it, building a skill takes about five minutes.
The live build: a second opinion skill
The first live demo was a skill Ilia wanted to create during the show: a “second opinion” skill. He wanted Claude to push back on his ideas, frameworks, and drafts from multiple perspectives instead of agreeing too easily.
That use case made sense immediately. Writers need it. Builders need it. Anyone making decisions with AI needs it. The point isn’t to make Claude negative. It’s to stop Claude from defaulting to “great idea” when you actually need useful pressure.
What to include in your prompt to the skill creator
When you invoke the skill creator, the cleanest results come from putting four things in the prompt:
