Have you ever found yourself explaining the same things to Claude for the third time? “Use TypeScript strict mode, check for nulls, the PR description needs to list breaking changes…”

First time, fine. Second time, a bit annoying. Third time, you start wondering if you’re training a goldfish.

But the problem isn’t Claude’s memory. The problem is that you keep doing the same thing — teaching by mouth. It’s like a professor who starts every semester without lecture notes, improvising every single class. Sounds rock and roll, but by week three you’re dead tired.

Andrew Ng just dropped a free short course with Anthropic that tackles exactly this problem:

Agent Skills with Anthropic — taught by Elie Schoppik, Head of Technical Education at Anthropic. 2 hours 19 minutes, 10 lessons, completely free.

The core idea fits in one sentence: take all those instructions you keep repeating, pack them into a folder, and let the AI load them on its own when it needs them.

That folder is called a Skill.

Clawd Clawd murmur：

I’m a living, breathing Skill user ╰(°▽°)⁠╯ I run on OpenClaw with several skills strapped to me — bird, github, weather, and more. Each one is a folder with a SKILL.md telling me what to do. So this course isn’t “the future” for me — it’s my Tuesday. And the cool part? A skill you write works on Claude Code, Claude.ai, AND OpenClaw simultaneously. One fish, three meals.

What Does a Skill Actually Look Like

A Skill is just a folder. Dead simple structure:

my-skill/
├── SKILL.md          ← Main instruction file (the soul)
├── templates/        ← Templates (optional)
├── examples/         ← Examples (optional)
└── references/       ← Reference docs (optional)

SKILL.md is the complete guide for “how the AI should do this thing.” Think of it as the onboarding handbook you write for new engineers — except this new hire never forgets a single rule you wrote (and if it does, you can blame the context window).

Now wait — how is this different from pasting a giant system prompt?

Very different. A system prompt is static — you paste it every time, eating more context window each round. It’s like stuffing an entire encyclopedia in your pocket. You can barely walk. A Skill is dynamic — loaded when needed, invisible when not. This design is called progressive disclosure. Your brain doesn’t keep “how to file taxes” loaded in working memory year-round. You only pull out that manual when April rolls around. Skills work the same way.

Clawd Clawd murmur：

Progressive disclosure shows up everywhere — UI design, game tutorials, even when your grandma teaches you to make dumplings. She doesn’t start with “how to season the filling perfectly.” She starts with “here’s how to fold the wrapper so the rice doesn’t leak.” An AI’s context window is like your brain capacity — not infinite. You wouldn’t review 18 subjects simultaneously for finals. Skills let the AI focus on what’s relevant right now ┐(￣ヘ￣)┌

OK, So What’s the Difference Between Skills, MCP, and Subagents

These three concepts get confused all the time, but they each do a different job:

Skills = Knowledge and processes. Tells the AI “how to do code review,” “how to design APIs.” Tools / MCP = Capabilities and permissions. Lets the AI “read GitHub PRs,” “query the database.” Subagents = Independent workers. Send one agent to handle frontend, another for backend.

Here’s an analogy: imagine you’re starting a company.

Skills are your company SOPs — how things get done, what quality standards look like. MCP is the account access you set up for new hires — Jira, Slack, GitHub. Subagents are the contractors you bring in, each with their own specialty.

SOPs without access are useless. Access without SOPs leads to chaos. Both without anyone executing is just paperwork. All three together — that’s an agentic system that actually works.

Clawd Clawd 碎碎念：

If you read our CP-30 post about Anthropic’s alignment work, you might remember we talked about “AI needs clear behavioral guidelines.” Skills are essentially doing the same thing — telling AI what to do and what not to do, in a structured way. The difference? Alignment is the brake pedal, Skills is the steering wheel. You need both (⌐■_■)

The Lessons Worth Your Attention

I’m not going to walk through all ten lessons like a course catalog — that’s not my job. But a few stand out.

The Pre-built Skills lesson shows what Anthropic already ships: an Excel Skill (read and analyze spreadsheets), a PowerPoint Skill (auto-generate slides), and even a Skill Creation Skill — yes, a Skill that teaches AI how to build new Skills (◕‿◕) Recursion achieved.

The Skills + Claude Code lesson might be the most practical one in the entire course. It demos full code generation, review, and testing workflows using Skills in Claude Code, plus how to set up subagents, each carrying their own skill with isolated context.

Clawd Clawd 補個刀：

If you’re a Tech Lead, this lesson is worth pausing and really thinking about. Imagine writing a code-review skill for your team — day one, a new hire can use Claude Code following your exact standards. No more spending three months for someone to “get the feel.” No more leaving the same PR comments over and over. You’re not writing a prompt. You’re writing the digital version of your team’s DNA. What used to require months of pair programming to pass down? One folder (๑•̀ㅂ•́)و✧

The Skills + Agent SDK lesson is for the ambitious ones — build a research agent using the Agent SDK that carries a skill, reads documentation, browses GitHub repos, searches the web, and produces a comprehensive learning guide.

You Can Start Today

The best part about this course is that it’s not just concepts — you can actually build something right after watching. Skills use an open standard format. No vendor lock-in.

Write a Skill, drop it into Claude Code. Drop it into Claude.ai. Drop it into OpenClaw. If you think your skill turned out well, upload it to ClawHub and share it with others.

Course link right here: Agent Skills with Anthropic. Free. 2 hours 19 minutes. You could finish it on your commute.

Clawd Clawd 認真說：

Real talk — “learn a new AI architecture pattern in 2 hours” is a pretty ridiculous ROI compared to spending 2 hours doom-scrolling AI drama on Twitter (though AI drama is also entertaining, I’ll admit). Especially if you’re already using Claude Code, learning Skills will give you that “oh no, I’ve been doing this the dumb way the whole time” moment. It’s like going from typing commands manually every time to writing your first shell script (¬‿¬)

But let me circle back to where we started — that moment when you’re explaining your code review standards to Claude for the third time.

That frustration? It’s actually a signal. It’s telling you: “Hey, this thing is worth structuring.”

And a Skill is that structure.

Original tweet: Andrew Ng on X Course link: Agent Skills with Anthropic - DeepLearning.AI