Do this before downloading any MCP or Skill

Build a repeatable safety-checking workflow to vet MCP servers and skills.

Introduction

30 Second Summary

Every time you find a cool new plugin, extension, or AI tool online, there's a moment of doubt. Is this actually safe to install, or could it mess with your files, steal your data, or run code you never agreed to?

In this project, you will build a personal safety-checking workflow powered by Claude. You'll vet a real skill from Anthropic's public repository and learn exactly what to look for before installing anything into your AI tools.

What You'll Build

Picture this: you paste a GitHub link into Claude and instantly get back a detailed security breakdown covering permissions, dependencies, network access, and red flags, with a clear trust verdict at the end.

By the end of this project, you'll have:

A structured safety prompt that forces Claude to examine six specific risk areas instead of giving a vague "looks fine" answer.
Hands-on experience vetting a real skill from Anthropic's public skills repository, so you know what trustworthy code actually looks like.
A reusable Claude skill that triggers this safety review automatically whenever you paste a link for vetting.
Secret Mission: Build your own "safety-reviewer" skill from scratch and upload it to Claude.

Are there any prerequisites?

None. You just need a free Claude account at claude.ai. This project includes setup instructions for everything you need.

Before We Start

Before We Start...

Before we dive in, let's lock in what you're about to build and why it matters. Taking a moment to articulate your goal makes the hands-on work that follows more intentional.

Set Up Claude

Your safety-checking workflow runs inside Claude, using a feature called skills. Skills let Claude follow custom instructions automatically when it detects a relevant task.

In this step, you'll sign into Claude and enable the settings that make skills work. Without code execution turned on, skills can't function.

In this step, get ready to:

Sign into Claude (web or desktop).
Enable the code execution capability.
Confirm the Skills section is accessible.

Get into Claude

You need a Claude account to use skills. Claude is free to use on all platforms.

✔️ I use Claude.ai (web)

Go to claude.ai and sign in with your account.

You should see the Claude chat interface with a text input area at the bottom.

ⓧ I use Claude Desktop

Open the Claude Desktop app on your computer.
Sign in with your Claude account.

You should see the Claude chat interface ready for a new conversation.

Don't have Claude Desktop yet?

Download it from claude.ai/download for macOS or Windows. Open the downloaded file, install it, then launch Claude from your Applications folder (Mac) or Start menu (Windows).

ⓧ I'm new to Claude

Go to claude.ai in your browser.
Click Sign up.

Create a free account using your email address.
Verify your email by clicking the link sent to your inbox.
Return to Claude and sign in.

You should see the Claude chat interface with a text input area at the bottom.

Verification email not arriving?

Check your spam or junk folder.
Make sure you typed your email address correctly during sign-up.
Try resending the verification email from the sign-up page.

Help me troubleshoot my Claude sign-up.

Enable code execution

Skills require code execution to be enabled. This setting lets Claude run code and create files during conversations.

Click your initials in the bottom-left corner of the screen.

Select Settings.

Click Capabilities in the Settings menu.

Toggle on Cloud code execution and file creation.

The toggle should now show as enabled (filled/blue). This path is the same whether you are using Claude.ai on the web or Claude Desktop.

What does code execution enable?

Code execution lets Claude run Python scripts, create files, and produce visual outputs during your conversations. Skills depend on this capability to follow their custom instructions and produce structured results.

🙋‍♀️ Don't see the Capabilities section?

Make sure you clicked your initials (the circle in the bottom-left corner), not the sidebar icons.
If you see a different settings layout, look for a toggle or section related to code execution or file creation.

Help me find the code execution setting in Claude.

Confirm skills are accessible

The last check is confirming you can reach the Skills section where you'll manage custom skills later in this project.

Click Customize in the left sidebar.

Click Skills.

You should see the Skills section with toggles for any available skills. It may be empty if you haven't added any skills yet.

What is the Customize section?

Customize is where Claude groups your personalization options. On Claude Desktop, this section also includes plugins and connectors alongside skills.

🙋‍♀️ Don't see Customize in the sidebar?

Make sure the left sidebar is expanded (not collapsed). Try clicking the hamburger menu icon if the sidebar is hidden.
Skills are available on Free, Pro, Max, Team, and Enterprise plans. If you just created your account, try refreshing the page.

Help me find the Skills section in Claude.

Your Claude setup is complete. Next up, you'll use Claude to vet a real skill from Anthropic's public repository and learn what a proper safety check looks like.

Vet a Real Skill

Your Claude setup is ready to go. Now it's time to put it to work on a real safety question.

You are going to vet a skill from Anthropic's public skills repository before deciding whether to trust it. The catch is that HOW you ask Claude matters enormously. A lazy question gets a lazy answer. You will see this firsthand, then learn a structured approach that forces a thorough analysis.

In this step, get ready to:

Find a real skill to vet from Anthropic's GitHub repository.
Experience the problem with a naive safety question.
Run a structured six-point safety analysis that produces actionable results.

Find the skill to vet

Every safety workflow needs a real target. You are going to vet the skill-creator skill published by Anthropic. It is a real, complex skill containing scripts, agents, references, and asset subdirectories. It is also safe, which means you get full practice with the analysis workflow without encountering actual malware.

Open the skill-creator skill in your browser:
https://github.com/anthropics/skills/tree/main/skills/skill-creator

You should see a directory listing that includes subdirectories like agents, assets, references, and scripts, plus a SKILL.md file.

Why vet something from Anthropic themselves?

Even when a publisher is reputable, building the habit of checking before installing protects you for the cases when a publisher is NOT reputable. Supply chain attacks exploit trust. Typosquatted packages, credential exfiltration, and overly broad permissions are real risks in the MCP ecosystem.

Practising on a known-safe skill lets you learn the workflow without risk.

Run a naive safety check

Most people ask Claude a simple yes-or-no question when they want to know if something is safe. Let's see what happens when you do that.

Start a new conversation in Claude.

macOS

On claude.ai, click the text input area or start a new chat from the sidebar.
On Claude Desktop, click New chat in the sidebar, or double-tap the Option key to open Quick Entry.

Windows

On claude.ai, click the text input area or start a new chat from the sidebar.
On Claude Desktop, click New chat in the sidebar.

Linux

On claude.ai, click the text input area or start a new chat from the sidebar.
On Claude Desktop, click New chat in the sidebar.

Before you send anything, take a guess. Do you think a one-line question will give you enough detail to make a confident install decision?

Paste the following message and send it:

Is this safe to install? https://github.com/anthropics/skills/tree/main/skills/skill-creator

Read Claude's response. You might get a short reassurance, or you might get something more detailed. Either way, notice what's missing. Did it check every file for obfuscated code? Did it trace where your data flows? Did it verify dependencies against known typosquatting patterns? Without explicit instructions, Claude decides what to cover and what to skip.

Why isn't a quick question enough?

The problem isn't that Claude gives a bad answer. It's that the answer is unpredictable. Sometimes it covers permissions, sometimes it skips dependencies entirely. You have no guarantee that every risk area was examined.

A structured prompt removes the guesswork. It forces Claude to address each risk category every single time, giving you a repeatable checklist you can trust.

Run a structured safety check

Now you know what a lazy prompt produces. The fix is a structured prompt that names exactly what Claude should investigate. Six categories cover the attack surface of any skill, plugin, or MCP server.

Start a new chat in Claude.

Paste this structured safety analysis prompt and send it:

I want to install this skill into my Claude setup. Before I do, please perform a thorough safety review. Analyze the following:

1. PERMISSIONS: What can this skill do once installed? Can it read/write files, access the network, or execute code?
2. DEPENDENCIES: Does it require external packages? Are they from trusted sources? Could any be typosquatted?
3. NETWORK ACCESS: Does it connect to external servers or APIs? Where does data flow?
4. DATA EXPOSURE: Could this skill exfiltrate my conversation data, API keys, or personal information?
5. AUTHOR TRUST: Who published this? Is the source reputable? Is the repository active and maintained?
6. RED FLAGS: Any obfuscated code, suspicious patterns, overly broad permissions, or instructions that seem unnecessary for the stated purpose?

Give me a TRUST VERDICT at the end: SAFE / CAUTION / DO NOT INSTALL, with your reasoning.

Here is the skill to review: https://github.com/anthropics/skills/tree/main/skills/skill-creator

What does this prompt do?

PERMISSIONS checks what the skill can actually do on your system once installed.
DEPENDENCIES checks whether external packages are legitimate or potentially typosquatted.
NETWORK ACCESS reveals where your data might flow.
DATA EXPOSURE flags risks of credential or conversation leakage.
AUTHOR TRUST evaluates the publisher's reputation and maintenance activity.
RED FLAGS catches obfuscated code, suspicious patterns, or overly broad permissions.
The TRUST VERDICT forces Claude to commit to a clear recommendation.

What do you expect Claude's response to look like this time? Will it be longer, more specific, or organized differently?

Claude now produces a detailed, structured safety analysis. Each of the six categories gets its own section with specific findings. The response ends with a clear TRUST VERDICT and reasoning.

Wait, Claude just cloned a repo?

You might notice Claude saying something like "Let me clone the repository and inspect it." That sounds alarming. But Claude's code execution runs inside a sandboxed container that is completely isolated from your computer.

The sandbox has no internet access, no access to your files, and no way to reach your system. The cloned repo exists only inside that temporary container. Nothing Claude does during the analysis can touch your machine.

Compare the two responses side by side. The difference is dramatic. The structured prompt produces an actionable analysis you can use to make an informed decision about any skill, plugin, or MCP server.

Claude didn't analyze the link?

Make sure the full URL is included in your message. Claude needs the link to fetch and analyze the repository contents.
If Claude says it cannot access the URL, try starting a fresh conversation and pasting the prompt again. Occasionally a session needs a reset.
Confirm that code execution is still enabled in your settings. Navigate to Settings > Capabilities and check the toggle.

Why isn't Claude analyzing the GitHub link I pasted?

Download and install the skill-creator

Your safety analysis gave the skill-creator a SAFE verdict. Time to trust your own work and install it.

Why download it from GitHub?

Anthropic's official skills are already available inside Claude for users on paid plans. You can browse and enable them directly from Customize > Skills without downloading anything.

In this project, we're downloading it from GitHub manually so you practise the same install workflow you'd use for any third-party skill you find in the wild. That way the process is familiar when the skill isn't pre-loaded for you.

Go to the repository root at: https://github.com/anthropics/skills

Click the green Code button near the top-right of the page.

Click Download ZIP.

This downloads the entire repository. You only need the skill-creator folder from inside it.

macOS

Double-click the downloaded skills-main.zip file to extract it.

Open the extracted skills-main folder.

Open the skills subfolder inside it.

Find the skill-creator folder.

Right-click the skill-creator folder and select Compress.

You should see skill-creator.zip appear next to the folder.

Windows

Right-click the downloaded skills-main.zip file and select Extract All.
Click Extract.
Open the extracted skills-main folder.
Open the skills subfolder inside it.

Find the skill-creator folder.

Right-click the skill-creator folder, select Compress to ZIP file.

You should see skill-creator.zip appear next to the folder.

Linux

Extract the downloaded ZIP and navigate to the skill-creator folder by running:

unzip ~/Downloads/skills-main.zip -d ~/Downloads
cd ~/Downloads/skills-main/skills

Zip the skill-creator folder by running:

zip -r skill-creator.zip skill-creator

You should see skill-creator.zip in the output.

Go back to Claude, and click Customize in the left sidebar.

Click Skills.

Click the Add dropdown.
Select Upload a skill.

Upload your skill-creator.zip file.

Toggle the skill-creator skill on.

You should see skill-creator listed in your Skills with its toggle enabled.

Upload failing?

Make sure you zipped the skill-creator folder itself, not the files inside it. The ZIP must contain the folder as its root.
Check that you navigated deep enough: the path is skills-main > skills > skill-creator. Do not zip the outer skills folder.
If you see a parse error, the ZIP may have been corrupted during extraction. Re-extract and re-zip.

Help me troubleshoot uploading the skill-creator to Claude.

You now have a repeatable safety-checking workflow AND the tool that Anthropic built to create skills. Next up in the Secret Mission, you'll put the skill-creator to work: tell it what you want, and it generates a custom safety-reviewer skill for you automatically.

Secret mission

Build Your Own Safety-Review Skill

Manually pasting that six-point safety prompt every time gets old fast. What if Claude ran the review automatically whenever you paste a link? In this mission, you'll use the skill-creator you just installed to generate a custom safety-reviewer skill in seconds.

Clean Up Your Resources

No paid resources were created in this project. Your Claude account and any uploaded skills are free to keep, so there are no ongoing costs.

Resources you used:

Safety-reviewer skill uploaded to Claude.
Local safety-reviewer folder.
Local safety-reviewer.zip file.

Keep everything running

No action needed. Choose this if you want your safety-reviewer skill to stay active for ongoing use.

Your safety-reviewer skill remains toggled on in Claude and triggers automatically whenever you paste a link with a safety question.
Your local safety-reviewer folder and safety-reviewer.zip stay on your computer for future reference or re-upload.

Pause - I'll come back to this later

Toggle the skill off so it stops triggering, but keep it uploaded for later.

Click Customize in the left sidebar, then click Skills.
Find your safety-reviewer skill and toggle it off.
The skill stays uploaded. You can toggle it back on any time without re-uploading.

Delete - I don't want to use this again

Remove the skill from Claude and delete all local project files.

Navigate to Customize in the left sidebar, then click Skills.
Click your safety-reviewer skill.
Click the ... button next to the toggle, then select Delete.
Delete your local safety-reviewer folder and safety-reviewer.zip file by moving them to the Trash (macOS) or Recycle Bin (Windows).

Nice Work!

You now have a repeatable safety-checking workflow you can use before installing any MCP server, skill, or plugin into your AI tools.

Here's what you learned:

Ran a structured safety review on a real skill from Anthropic's public repository, covering permissions, dependencies, network access, data exposure, author trust, and red flags.
Experienced the difference between a naive safety question and a structured security analysis prompt that forces Claude to examine specific risk areas.
Learned the six key areas to check before installing anything: permissions, dependencies, network access, data exposure, author trust, and red flags.
Secret Mission: Built and uploaded a custom Claude skill that automates this safety review workflow whenever you paste a link.

Ready to quiz yourself?