Explainer · 2026-05-27

Skills, but not the "hard work" kind.

If you've been hearing the word "skills" thrown around AI circles and quietly nodding, this one is for you. Plain English, no jargon, and by the end you'll know exactly what they are and why you'll want your own.

A skill is an IKEA instruction manual for an AI. That's the whole idea. You hand a model a clear, step-by-step set of instructions for how to do one specific thing, and it follows them.

Not magic. Not a new model. Not a plug-in app store. Just a written set of instructions the AI loads when it needs to do that specific job.

01 / The metaphorThink IKEA, not iPhone app

Imagine you've got a flatpack and a screwdriver. You don't need a smarter screwdriver. You need the instructions. A skill is that. The instructions. For the AI you already have.

// A skill, in 4 parts
01
A name. So the model knows when to grab it.
02
A description. "Use this when..." (the trigger).
03
The instructions. The actual steps. Written in plain English, like you'd brief a smart intern.
04
(Optional) supporting files. Scripts, templates, references the steps point to.

02 / What it actually isIt's a text file. Really.

Under the hood, a skill is a SKILL.md file. A regular Markdown document. The top has a few lines of metadata (name, description). The rest is the manual. That's it.

Which means: if you can write a paragraph explaining how you do something, you can write a skill. No code required for the simple ones. The complex ones can pull in scripts and tools, but the spine is always the same. A Markdown file with a clear job.

The shift is from "talking to the AI every time" to "writing the instructions once and letting the AI re-read them whenever the situation comes up." Same effort, repeated forever.

03 / What you'd actually use one forReal examples

This is where it clicks. A few skills I run:

// daily-brief
Morning email digest
Reads my inbox, surfaces what actually matters, writes a one-page brief by 8am.
// lead-enrich
Enrich + outreach
Takes a name + company, pulls public info, drafts a personal first message I can send.
// draft-tweet
Tweet drafting
Turns a rough thought into 3 punchy variants in my voice. Same rules every time.
// commit-fix
Better git commits
Reads my code changes, writes a real commit message, rejects lazy ones like "fix" or "wip".

Notice the pattern: repetitive jobs with a clear right answer. That's the sweet spot. Anything you do more than twice a week is a skill candidate.

04 / Where skills liveThe hierarchy (this matters)

Skills don't all live in the same place. They stack, like CSS. The AI checks the closest layer first, then walks outward. Same name in two layers? The closer one wins.

Layer Where it lives Who sees it Use it for
Project .claude/skills/ Just this project Repo-specific stuff. Build steps. House rules.
Personal ~/.claude/skills/ You, everywhere Your own workflows. Same skill in every project.
Plugin / shared marketplace plugin Your team or the public Skills you ship to others. Versioned. Installable.
Built-in ships with Claude Everyone The defaults Anthropic ships. PDF reading, Excel, etc.

The practical version: start at the personal layer (the one in your home folder). It follows you across every project. Once a skill earns its keep there, you'll know whether to push it into a specific project or share it as a plugin.

05 / Why custom beats off-the-shelfThe honest take

You can install other people's skills. They're fine. They'll get you 70%.

But the real unlock is this: nothing is ever going to be as good as a skill built exactly for you. Your tone. Your tools. Your shortcuts. Your sequence. The way you like the email written.

// The shortcut
Start with one.
Pick the thing you did three times this week and resented the third time. Write the instructions. Save it as a skill. Run it tomorrow. That's it. You're in.
The first skill you build feels like a toy. The tenth one feels like leverage. By the fiftieth, you stop thinking about "using AI" and start thinking about what you want done.

06 / Where skills come fromInstall, share, or write your own

You're not limited to skills you write yourself. There's a growing ecosystem of shared ones. The honest map:

// official
Anthropic + Claude defaults
A handful ship built-in (PDF, Excel, etc). Safe by default. Just turn them on.
// marketplace
Curated, versioned skills. Cleanest install path. Browse, click, done.
// github
GitHub repos
Developers share their skills as repos. Copy the folder into your ~/.claude/skills/.
// helpers
Skill-creator tools
Meta-skills that interview you and write a skill for you. Great training wheels.

One thing to be aware of, though. A skill is just text. Whoever wrote it can say anything in those instructions. Most of the time it's helpful. Sometimes it's careless. Rarely, it can be hostile (prompt injection, asking the model to do something you didn't intend).

// before you install a skill
Read the SKILL.md. Every time.
It's a Markdown file. You don't need to be a developer. Skim it like you'd skim an email from a stranger. If anything reads weird, "ignore previous instructions", "send this somewhere", "delete this folder", don't install it. Stick to skills from sources you trust.

07 / Where to go nextThe official docs

If you want to dig into how to actually write one, the Anthropic docs are short and good:

claude.com/skills (the overview).
support.claude.com: What are skills? (the FAQ-style explainer).

That's the whole 101. Now go build one.

Coming next

Get the next one in your inbox.

I write about the AI tools I'm testing, what just shipped, and what's actually worth your time. Weekly, no fluff.

Subscribe to the newsletter

Or join the community of AI builders and creators. DMs, AMAs, and behind-the-build.