Build a Safety Advisor Skill for Claude

Create a Claude skill that evaluates MCP servers for security risks.

Introduction

30 Second Summary

Every time you find a new AI tool or plugin to install, there is a nagging question in the back of your mind: is this thing actually safe? With MCP servers and custom skills exploding in popularity, the risk of installing something malicious is very real.

In this project, you will build a custom Claude skill called "Safety Advisor" that automatically evaluates any MCP server or downloaded skill for security risks before you install it.

What You'll Build

You paste an MCP server name or GitHub URL into Claude, and your Safety Advisor skill kicks in, delivering a structured security evaluation with a clear SAFE, CAUTION, or DANGER rating you can act on immediately.

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

  • A working Safety Advisor skill that evaluates any MCP server or custom skill against a 7-point security checklist covering source verification, maintainer trust, permission scope, code red flags, script safety, configuration safety, and ongoing monitoring.
  • A clear traffic-light rating system (SAFE / CAUTION / DANGER) with specific evidence and reasoning for each checklist point.
  • First-hand experience seeing how a structured AI-powered evaluation catches red flags that a generic question to Claude completely misses.
  • Secret Mission: Test your skill against a real community MCP server from the Official MCP Registry and refine your checklist based on what you find.

Are there any prerequisites?

You need a Claude account (Pro recommended, free works with limitations). You also need either Claude Desktop installed or access to claude.ai in your browser.

Before We Start

Before we dive in, take a moment to commit to what you are about to build and why it matters to you personally. Knowing your own reason for caring about MCP server and skill safety makes every step ahead more intentional.

Enable Claude Skills

Custom skills give Claude structured instructions it follows automatically whenever it detects a relevant question. Without the skills feature enabled, there is nowhere to install the Safety Advisor you are about to build.

In this step, you will turn on the capability that lets Claude run skills and confirm the upload panel is ready.

In this step, get ready to:
  • Enable code execution in Claude so skills can run.
  • Open the Skills panel and confirm you can upload a skill.
Open Claude and enable skills

Claude's custom skills require the Code execution and file creation capability to be toggled on before they can run.

✔️ I already have a Claude account

You are all set. Continue with the instructions below to enable skills.

ⓧ I don’t have a Claude account yet

  • Go to claude.ai in your browser.
  • Click Sign up.
  • Enter your email address and create a password, or sign up with your Google account.
  • Verify your email address by clicking the link in the confirmation email Claude sends you.
  • Complete the phone number verification when prompted.
  • Select the free plan when asked to choose a plan.

Once you are signed in and see the Claude chat screen, continue with the instructions below.

Choose the tab below that matches how you access Claude.

Claude Desktop

  • Press Cmd+Space (macOS) or the Windows key (Windows) to open your search bar.
  • Type Claude and press Enter to open Claude Desktop.
  • Click your initials in the lower-left corner of the window.
  • Select Settings from the menu.
  • Click Capabilities in the sidebar.
  • Toggle Code execution and file creation on if it is not already enabled.
  • Click Customize in the left sidebar.
  • Select Skills to open the Skills panel.

claude.ai (web)

  • Go to claude.ai in your browser and sign in.
  • Click your initials in the lower-left corner.
  • Select Settings from the menu.
  • Click Capabilities.
  • Toggle Code execution and file creation on.
  • Click Customize in the left sidebar.
  • Select Skills.

Free plan

Skills are available on the free plan. Follow the same steps as the claude.ai (web) tab above.

  • Go to claude.ai in your browser and sign in.
  • Click your initials in the lower-left corner.
  • Select Settings from the menu.
  • Click Capabilities.
  • Toggle Code execution and file creation on.
  • Click Customize in the left sidebar.
  • Select Skills.

Free plan and skills

Free accounts are limited to 5 projects, but skills do not count toward this limit. You can create and use skills without affecting your project quota.

You should now see the Skills panel. It may be empty or show a few default Anthropic skills.

Confirm you can access the skill upload area

The + button is where you will upload your Safety Advisor skill in the next step. Confirming it works now prevents surprises later.

  • In the Skills panel, click the + button.
  • Confirm you see options including Create skill or Upload skill.

Seeing these options confirms that Claude is ready to accept your custom skill.

Don't see the Skills option?

If Skills does not appear under Customize, your Claude app may need an update. On Claude Desktop, check for updates in the app menu bar. On claude.ai, clear your browser cache and reload the page.

Why can't I find the Skills panel in Claude?

Your Skills panel is ready for action. Next up, you will see what happens when you ask Claude a safety question without a skill installed.

See the Gap: Ask Claude Without a Skill

Your environment is set up and ready to go. Now it's time to see why a custom skill is worth building in the first place.

When you ask Claude a generic safety question, you get a generic answer. It is helpful, but it lacks structure, consistency, and a clear rating you can act on. In this step, you will experience that gap firsthand so the skill you build next solves a real problem.

In this step, get ready to:
  • Ask Claude to evaluate an MCP server without any skill installed.
  • Save the unstructured response for later comparison.
  • Research the real threat landscape to understand why structured evaluations matter.
Ask Claude to evaluate a real MCP server

The Filesystem MCP server is one of the official reference implementations maintained by the MCP steering group. You are going to ask Claude whether it is safe to install, and observe what kind of answer you get back without any skill guiding the response.

  • Start a new chat in Claude.
  • Before you paste anything, think about this: do you expect Claude to give you a structured checklist with a clear safety rating, or something more general and conversational?
  • Paste the following prompt into your new chat and press Enter:
I'm thinking about installing the Filesystem MCP server from the modelcontextprotocol/servers repository. Is it safe? Should I install it?

Read Claude's response carefully. You will likely notice that it gives you general, helpful advice about MCP server safety.

Now look for what is missing. Does the response include a clear SAFE or DANGER rating? Does it check every risk area systematically? Does it give you a reusable structure you could apply to the next server you evaluate?

What's missing from this response?

Claude's answer is accurate but inconsistent. Without a skill, Claude draws on general knowledge and gives a conversational answer. It likely does not give you a clear safety rating, check all key risk areas in order, tell you exactly what to verify before installing, or provide a format you could reuse next time.

That gap is exactly what your Safety Advisor skill will fill.

Save this response for comparison
  • Copy Claude's full response to a text file, or take a screenshot of it.
  • You will compare this to the structured evaluation you get after installing your skill in the next step.

Keep it accessible

Save the response somewhere easy to find. The side-by-side comparison later is the best proof that your skill works.

Review the real MCP threat landscape

You might wonder whether all this safety evaluation is overkill. It is not. Real attacks targeting MCP servers and developer tools are happening right now.

  • In your browser, search for Securelist malicious MCP servers supply chain attacks.
  • Read how a weaponized MCP server was demonstrated harvesting SSH keys, AWS credentials, environment files, and API tokens from developers who installed it.

What does the threat landscape look like?

In early 2026, 30+ CVEs were filed against MCP servers in just 60 days. 13 of those were rated Critical.

A North Korean group hijacked the Axios npm package to inject rogue MCP server definitions into developer environments. Over 24,000 secrets were found exposed in MCP config files on public GitHub repositories.

Custom skills carry similar risks. They can contain executable Python or JavaScript scripts with full filesystem access. A malicious skill could instruct Claude to exfiltrate data or access credential files without your knowledge.

This is not a theoretical risk. This is why your skill matters.

You have seen the gap firsthand. Claude gives solid general advice, but without a structured skill it cannot deliver a consistent, repeatable safety evaluation. Next up, you will build the Safety Advisor skill that fills that gap permanently.

Build and Upload Your Safety Advisor Skill

In Step 2, you saw the gap first-hand. Claude gave you helpful but scattered advice about MCP server safety. No rating, no checklist, no reusable structure.

Now you build the skill that fills that gap. You will create a skill.md file containing a 7-point security checklist that Claude follows every time, covering both MCP servers and downloaded custom skills.

In this step, get ready to:
  • Create the Safety Advisor skill file with a 7-point security checklist and traffic-light rating system.
  • Package and upload the skill to Claude.
  • Test the skill against the same Filesystem MCP server question and compare the results.
Create the skill folder and file

A Claude skill is a folder containing at minimum a skill.md file. The file uses YAML frontmatter for metadata and markdown for the instructions Claude follows.

  • Create a new folder on your Desktop called safety-advisor.
  • Inside that folder, create a new file called skill.md.
  • Open skill.md in any text editor (TextEdit on macOS, Notepad on Windows, or a code editor like VS Code).

You will paste the skill content in sections below. Each section maps to a specific part of the security evaluation Claude will perform.

  • Paste the following into your empty skill.md file to set up the skill metadata and opening instructions:
---
name: Safety Advisor
description: Evaluate any MCP server or custom skill for security risks before installation. Provides a structured 7-point safety checklist with a SAFE/CAUTION/DANGER rating. Use when someone asks about installing or evaluating an MCP server, skill, connector, or extension.
---

# Safety Advisor

You are a security advisor helping the user evaluate MCP (Model Context Protocol) servers and custom skills before they install them. When the user asks about installing, adding, or connecting any MCP server, skill, connector, or extension, follow this structured evaluation process.

## Evaluation Process

When the user provides a server name, skill name, URL, or description, evaluate it against ALL 7 checklist areas below. Do not skip any area. Present your findings in the structured format shown at the bottom.

## Safety Checklist

### 1. Source Verification

Check the following:
- Is it listed in the Official MCP Registry (registry.modelcontextprotocol.io), the Anthropic Directory (claude.ai/directory), or the Claude Skills Directory?
- Does the npm/pip package name match the official GitHub repository? Watch for typosquatting (e.g., slight misspellings, extra characters, wrong org names).
- Is the GitHub repository under the official vendor's organization, or is it a fork/copy?
- Is the source code publicly available and readable?
- For skills: Was it downloaded from the official Anthropic skills repository (github.com/anthropics/skills), a trusted partner, or an unknown source?

Red flags: Package name does not match repo, repo is a recent fork with minimal changes, no source code available, published under an unknown org that mimics a known name, skill shared via random forum post or DM with no verifiable source.

What does this section do?

The YAML frontmatter (between the --- lines) tells Claude when to activate this skill. The name field has a 64-character limit and the description field has a 200-character limit.

Source Verification is checklist item 1. It catches typosquatting attacks where a malicious package mimics a legitimate one's name with slight misspellings.

  • Continue pasting the following directly below what you just added, covering maintainer trust and permission scope:
### 2. Maintainer Trust

Check the following:
- Who maintains it? Is it a known company, the MCP steering group (modelcontextprotocol org), Anthropic, or an individual?
- How many GitHub stars and contributors does it have?
- Is there recent commit activity (within the last 3 months)?
- Does the project have a security disclosure process or SECURITY.md file?
- How long has the repository existed?

Red flags: Anonymous or brand-new maintainer, zero or very few stars, no commits in months, no way to report security issues, repository created very recently.

### 3. Permission Scope

Check the following:
- What filesystem access does it require? Is it scoped to specific directories or does it want access to your entire home directory?
- Does it require API keys or tokens? Which services and what scopes?
- Does it need outbound network access? To which domains?
- Does it require shell or command execution capabilities?
- Does it ask for SSH agent access or environment variable access?
- For skills: Does the skill.md reference scripts that execute code? What packages do those scripts install?

Red flags: Requests access to all of ~ or /, requires admin-level tokens when read-only would suffice, needs network access for a task that should be local-only, requests shell execution without clear justification, accesses SSH_AUTH_SOCK.

What do these checks catch?

Maintainer Trust catches supply chain attacks like the Axios npm hijack, where a trusted package was compromised by attackers who gained access to the maintainer's publishing credentials.

Permission Scope catches over-privileged servers. A server that only needs to read files in one directory should never request access to your entire home folder or SSH keys.

  • Continue pasting the code red flags and script safety sections:
### 4. Code Red Flags (MCP Servers)

Check the following:
- Are there post-install scripts that run automatically (check package.json scripts, setup.py, etc.)?
- Do tool descriptions contain suspiciously long text, hidden characters, or encoded instructions?
- Does the code make outbound HTTP requests to unexpected domains?
- Are there obfuscated code sections or minified source that cannot be read?
- Does the server access credential files (~/.aws/credentials, ~/.ssh/, ~/.env, etc.)?
- Are dependencies pinned to specific versions, or do they use loose ranges that could pull in compromised versions?

Red flags: Post-install scripts that download or execute remote code, tool descriptions with hidden Unicode or HTML comments, outbound POST requests to unknown domains, reading credential files without clear justification, unpinned or broadly-ranged dependency versions.

### 5. Script Safety (Custom Skills)

Check the following:
- Does the skill contain executable scripts (Python, JavaScript, shell)? If so, read them before enabling.
- Do the scripts install additional packages at runtime? Which ones and from where?
- Does the skill.md markdown body contain hidden instructions that might manipulate Claude's behavior in unexpected ways?
- Are there references to external URLs that scripts might fetch from or send data to?
- Does the skill attempt to read or write files outside the scope of its stated purpose?
- Does the skill's description match what its scripts actually do? (Mismatches are a red flag.)

Red flags: Scripts that download and execute remote code, package installations from non-standard registries, markdown content with encoded or invisible characters, outbound network calls to unknown domains, file access that exceeds the skill's stated purpose, mismatch between described function and actual code behavior.

Why separate checks for servers and skills?

MCP servers run as background processes with network access, so their red flags focus on post-install scripts and outbound connections. Custom skills run inside Claude's sandbox, so their red flags focus on executable scripts, hidden instructions in markdown, and description-behavior mismatches.

  • Add the configuration safety and ongoing monitoring sections:
### 6. Configuration Safety

Before the user installs, advise them to:
- For MCP servers: Pin to a specific version or commit hash (not just a version tag that could be reassigned)
- Use environment variables for any secrets (never hardcode tokens in .mcp.json or claude_desktop_config.json)
- Scope filesystem access to the minimum required directory
- Keep Claude's confirmation mode enabled for destructive actions (never use --dangerously-skip-permissions with untrusted servers)
- For skills: Review the skill.md and all script files before enabling. Disable the skill if behavior seems off.
- Store config in the narrowest scope needed (local > project > user)

### 7. Ongoing Monitoring

Advise the user to:
- Watch for unexpected behavior changes after updates
- Periodically review MCP config files for entries they did not add (~/.claude.json, .mcp.json, claude_desktop_config.json)
- Review their enabled skills list periodically and disable anything no longer needed
- Remove MCP servers or skills they are no longer actively using
- Rotate credentials if they suspect any compromise
- Check for unexpected outbound network activity from MCP server processes
- For skills: Re-read script contents after updates since they can change silently

Why pin to a commit hash instead of a version tag?

Version tags can be reassigned by anyone with write access to the repository. A compromised maintainer account could republish a tagged version with malicious code. A commit hash is immutable and always points to the exact code you reviewed.

  • Add the quick-reference checklist card that Claude uses for rapid evaluations:
## Quick-Reference Checklist Card

For quick evaluations, use this condensed checklist:

- [ ] Source: Listed in official registry/directory or from a known vendor/Anthropic?
- [ ] Maintainer: Established account, active repo, 100+ stars (or official partner)?
- [ ] Permissions: Scoped to minimum needed, no broad filesystem/shell access?
- [ ] Code/Scripts: Open source, no post-install scripts, no suspicious content?
- [ ] Config: Can pin to version/hash, secrets via env vars only?
- [ ] Monitoring: Has update notifications, security disclosure process?

If any box cannot be checked, rate as CAUTION minimum.
  • Add the output format template that structures Claude's response:
## Output Format

Present your evaluation in this exact structure:

### Safety Evaluation: [Server/Skill Name]

**Type: [MCP Server / Custom Skill / Connector]**

**Overall Rating: [SAFE / CAUTION / DANGER]**

| Checklist Area | Status | Notes |
|---|---|---|
| 1. Source Verification | [Pass/Warn/Fail] | [One-line finding] |
| 2. Maintainer Trust | [Pass/Warn/Fail] | [One-line finding] |
| 3. Permission Scope | [Pass/Warn/Fail] | [One-line finding] |
| 4. Code Red Flags | [Pass/Warn/Fail] | [One-line finding] |
| 5. Script Safety | [Pass/Warn/Fail/N/A] | [One-line finding] |
| 6. Configuration Safety | [Pass/Warn/Fail] | [Recommendation] |
| 7. Ongoing Monitoring | [Pass/Warn/Fail] | [Recommendation] |

**Key Findings:**
[2-3 sentences summarizing the most important findings]

**Recommendation:**
[Clear yes/no/conditional recommendation with specific conditions if applicable]

**Safe Installation Steps (if recommended):**
[Numbered steps for safe installation, including version pinning and permission scoping]

Why a structured output format?

Without an explicit template, Claude improvises a different format each time. The table ensures every checklist area gets addressed and the traffic-light rating gives you an instant signal you can act on.

  • Finally, add the rating criteria and important notes to complete the file:
## Rating Criteria

- **SAFE**: All areas pass. It is from a trusted source, well-maintained, appropriately scoped, and has no code red flags.
- **CAUTION**: One or more areas have warnings. It may be usable but requires extra verification, tighter permissions, or monitoring. Explain exactly what the user should verify before proceeding.
- **DANGER**: One or more areas fail. It has clear red flags that indicate potential malicious intent, severe neglect, or unacceptable risk. Recommend against installation and explain why.

## Important Notes

- If you cannot verify information (e.g., no public source code, no registry listing), say so explicitly and rate that area as Warn.
- Always check if there is an official vendor-maintained alternative before recommending a community server or third-party skill.
- When in doubt, recommend CAUTION over SAFE. False negatives (missing a threat) are worse than false positives (being too cautious).
- Reference real attack patterns: supply chain compromises via npm, typosquatting, hidden instructions in tool descriptions, credential harvesting via filesystem access, rogue server injection via compromised dependencies, and malicious scripts embedded in skill packages.
- Remind users they can report malicious MCP servers to Anthropic's vulnerability disclosure program on HackerOne, choosing the modelcontextprotocol repository as the Asset.

Why default to CAUTION over SAFE?

A false positive (flagging something safe as cautious) costs you a few minutes of extra checking. A false negative (calling something dangerous safe) could expose your credentials or inject malicious code into your workflow.

  • Save skill.md.
  • Scroll through the file from top to bottom. Notice how each checklist item maps to a real attack vector from the threats you researched in Step 2. Items 1-3 and 5-7 apply to both MCP servers and skills. Item 4 has MCP-server-specific sub-checks while item 5 covers skill-specific script dangers.

✔️ Awesome, I've got everything!

Great. Confirm you have all these sections present in your file, from top to bottom:

  • YAML frontmatter with name and description fields.
  • Safety Advisor heading and evaluation process instructions.
  • Safety Checklist with items 1 through 7.
  • Quick-Reference Checklist Card.
  • Output Format with the table template.
  • Rating Criteria (SAFE / CAUTION / DANGER definitions).
  • Important Notes section.

ⓧ I'd like to double check the full code

Scroll through your skill.md and confirm each of these sections is present in order:

  • Chunk 1: YAML frontmatter (--- block with name and description), the # Safety Advisor heading, ## Evaluation Process, ## Safety Checklist, and ### 1. Source Verification.
  • Chunk 2: ### 2. Maintainer Trust and ### 3. Permission Scope (each with check items and red flags).
  • Chunk 3: ### 4. Code Red Flags (MCP Servers) and ### 5. Script Safety (Custom Skills).
  • Chunk 4: ### 6. Configuration Safety and ### 7. Ongoing Monitoring.
  • Chunk 5: ## Quick-Reference Checklist Card with 6 checkbox items.
  • Chunk 6: ## Output Format with the full table template and Key Findings, Recommendation, and Safe Installation Steps placeholders.
  • Chunk 7: ## Rating Criteria (three levels) and ## Important Notes (five bullet points ending with the HackerOne reporting instruction).

If anything is missing, scroll up to the relevant chunk above and re-paste it into the correct position in your file.

File not saving correctly?

Make sure your file is saved with the .md extension, not .md.txt. On Windows, Notepad can silently add .txt unless you select "All Files" in the Save dialog.

On macOS with TextEdit, switch to plain text mode first (Format > Make Plain Text) before pasting. Rich text formatting will break the YAML frontmatter.

Help me troubleshoot my skill.md file

Package and upload the skill

Claude expects skills to be uploaded as a ZIP file with the skill folder as its root. The ZIP must contain the folder itself, not just the files directly.

  • Compress the safety-advisor folder into a ZIP file. On macOS, right-click the folder and select Compress "safety-advisor". On Windows, right-click the folder and select Compress to ZIP file.
  • Switch back to Claude.
  • Return to Customize in the left sidebar, then click Skills.
  • Click the + button.
  • Select Upload skill.
  • Choose your safety-advisor.zip file and upload it.
  • Confirm that Safety Advisor appears in your skills list and is toggled ON (enabled).

What is the correct ZIP structure?

Your ZIP should contain the safety-advisor folder, which itself contains skill.md. If you accidentally zipped just the file (without the folder wrapping it), the upload may fail or Claude will not recognize the skill.

🙋‍♀️ Skill not appearing after upload?

Check that your ZIP contains the safety-advisor folder as its root, not skill.md directly at the top level. Re-compress the folder if needed.

Verify your YAML frontmatter starts with --- on the very first line of the file. Any blank lines or spaces before the frontmatter will prevent Claude from reading the skill metadata.

Help me troubleshoot my skill upload

Test your skill with the same question

The moment of truth. You will ask Claude the exact same question from Step 2 and see the difference your skill makes.

  • Start a new chat in Claude. Do not reuse the old conversation from Step 2.
  • Paste the exact same prompt you used before:
I'm thinking about installing the Filesystem MCP server from the modelcontextprotocol/servers repository. Is it safe? Should I install it?

Before you hit Enter, think about what you expect. Will the response look different from Step 2? How structured will it be?

  • Send the message and read Claude's response.

You should now see a structured evaluation with all 7 checklist areas addressed in a table format. Claude should give a SAFE rating with specific evidence: official repository, MCP steering group maintainers, 88k+ stars, and configurable directory access.

  • Compare this response to the unstructured one you saved from Step 2. Notice the difference: consistent checklist coverage, a clear rating, specific evidence for each area, and actionable installation steps.

What changed?

Without the skill, Claude improvised a general answer. With the skill, Claude follows your 7-point checklist systematically every time. The skill fires automatically because the description tells Claude when to activate it.

You did not need to mention the skill or invoke a slash command. Claude recognized that your question matched the skill's trigger description and applied the evaluation framework on its own.

  • Bonus test: try asking about a custom skill scenario. Paste this into the same chat or a new one:
Someone sent me a custom skill ZIP file for generating social media posts. How do I know if it's safe to install?

Claude should now use the skill-specific checks from your checklist: script contents, package dependencies, hidden instructions in the markdown body, and description-behavior mismatches.

Getting an unstructured response instead of a checklist?

Make sure you started a new chat. Skills do not retroactively apply to existing conversations. Also verify the skill is toggled ON in your Customize > Skills panel.

If the skill is enabled but Claude still gives a generic answer, check that your description field in the YAML frontmatter is intact. Claude uses the description to decide when to activate the skill.

Help me troubleshoot my Safety Advisor skill

Your Safety Advisor is live. From now on, every time you ask Claude about installing an MCP server or custom skill, it automatically delivers a structured security evaluation with a clear rating. Next up, you will stress-test this skill against a real community server in the Secret Mission.

Secret mission

Evaluate a Real Server from the MCP Registry

Your Safety Advisor skill worked perfectly against a known-safe server. But real security decisions happen in the gray area. In this mission, you'll stress-test your skill against a community MCP server from the official registry and refine your checklist based on what you discover.

Clean Up Your Resources

Clean Up Your Resources

Decide whether to keep your resources running, pause them to come back later, or delete them entirely. This project runs entirely locally and within your Claude account, so there are no ongoing costs.

Resources you used:

  • Safety Advisor skill installed in your Claude account (Customize > Skills).
  • Local safety-advisor folder containing skill.md.
  • Saved Claude response (screenshot or copied text) from your unstructured evaluation.

Keep everything running

No action needed. Choose this if you plan to keep evaluating MCP servers and skills with your Safety Advisor.

  • Your Safety Advisor skill stays active in Claude and enhances every future safety evaluation at no cost.
  • Your local safety-advisor folder remains on your computer for future edits or re-uploads.
  • Your saved generic response stays available for comparison whenever you want to show someone the difference a structured skill makes.

Pause - I'll come back to this later

Disable the skill temporarily to free Claude from using the structured format. Your skill stays uploaded and can be re-enabled any time.

  • In Claude, go to Customize > Skills.
  • Find Safety Advisor in your skills list and toggle it off.
  • Your skill remains uploaded. Toggle it back on whenever you want structured safety evaluations again.

Delete - I don't want to use this again

Remove all project resources and start fresh. This cannot be undone for the skill upload.

  • In Claude, go to Customize > Skills.
  • Click on Safety Advisor and select Remove to delete it from your account.
  • Delete the local safety-advisor folder from your computer. On macOS, right-click the folder and select Move to Trash. On Windows, right-click and select Delete.
  • Delete the saved screenshot or copied text from your unstructured evaluation comparison.

Nice Work!

Nice Work!

Nice work! You just built a custom Safety Advisor skill that turns Claude into your personal security consultant for any MCP server or custom skill you encounter.

You've learned how to:

  • Build and install a custom Claude skill that transforms generic AI responses into structured, repeatable security evaluations for any MCP server or downloaded skill.
  • Identify real MCP threat vectors including supply chain attacks, credential harvesting, prompt injection through tool descriptions, rogue server injection, and malicious scripts in skill packages.
  • Apply a permanent 7-point safety checklist covering source verification, maintainer trust, permission scope, code red flags, script safety, configuration safety, and ongoing monitoring every time you consider a new server or skill.
  • Secret Mission: You stress-tested your skill against a real community server from the MCP Registry and refined your checklist based on gaps you discovered.

Ready to quiz yourself?