How to check an AI skill before your HR team uses it
Before HR uses an AI skill, GPT, or agent, check who made it, read its source, find where employee data goes, and compare its claims with its own docs.
AI HR Skills Editorial TeamBefore your HR team uses an AI skill, check four things: who made it, whether you can read its source, what employee data it touches and where it goes, and whether its claims match its own documentation. You can run all four checks from the tool’s own pages. Do it before anyone types in real employee data.
These tools don’t always arrive through a formal buying process. A colleague shares a SKILL.md file for Claude, someone finds a custom GPT that writes job ads, or IT turns on a new Copilot agent. Each one can read what you type, and some can send it somewhere else.
Key takeaways
- Look for a named author you could contact, because an anonymous tool leaves no one to ask.
- Read the source, not just the description: Anthropic’s documentation says to audit every file in a skill from an unknown source.
- Find out where employee data goes: OpenAI says it doesn’t audit or control how outside services connected to a GPT use or store data.
- If a tool’s claims and its own documentation don’t match, believe the documentation.
Check 1: Who made it
Start with the author. A tool you’d trust with employee data should name a person or company. You should be able to reach them with a question, a bug report, or a data request.
A named author doesn’t make a tool safe. It does make someone accountable for it. Say a skill arrives as a file with no name on it, or a GPT sits under a handle with no other trace. Then you have no one to ask what it does with your data.
Check 2: A source you can read
Next, open the tool itself. What you can read depends on the type of tool:
- A Claude skill is a folder with a SKILL.md file of plain-text instructions, sometimes with scripts. You can read every file.
- A custom GPT shows you its description and conversation starters. The person who built it sets its instructions and any outside services it connects to. OpenAI plans to retire custom GPTs on December 11, 2026, and move them into plugins, so ask what replaces one before your team adopts it.
- A Microsoft 365 Copilot agent is managed by your admins. According to Microsoft’s documentation, admins can see the permissions and data access an agent needs in the Microsoft 365 admin center. They can also read its terms of use and privacy statement, and choose which agents are allowed.
For skills, the maker’s own advice is blunt. Anthropic’s documentation says to use skills only from trusted sources. If you must use one from an unknown source, it says to audit every bundled file first.
It also warns that a malicious skill can make Claude call tools or run code in ways that don’t match its stated purpose. Its advice is to treat a skill like any software you install.
Check 3: What employee data it touches and where it goes
This is the check that matters most for HR. Start with two questions.
What will people put into it? SHRM’s AI tool evaluation checklist asks whether a tool will collect personal identifying information, performance and compensation data, or health information. A tool that drafts job ads and a tool that summarizes performance notes carry different levels of risk.
Where does that data go after you type it? Some tools send part of what you type to another service. Here’s what the makers say:
- OpenAI’s help center says that when a GPT uses apps or external APIs, relevant parts of your input may be sent to that outside service. It adds that OpenAI doesn’t audit or control how those services use or store your data.
- Microsoft’s extensibility guidance says services hosted outside Microsoft 365 stay responsible for their own permissions, privacy, and compliance. It says organizations should only turn on declarative agents that use trusted knowledge sources and trusted APIs.
- Anthropic’s enterprise checklist tells reviewers to search a skill’s files for network calls and external URLs.
SHRM’s checklist adds a question that’s easy to skip. Does the tool run on its own AI model, or on another company’s? If it’s another company’s, ask how much of your data that company can access.
Check 4: Do its claims match its own documentation?
A tool’s description is a pitch. Its documentation is where the details live. Read both and compare them.
Watch for capability words the documentation doesn’t back up. “Anonymous,” “secure,” “compliant,” and “works in Copilot” all need a source behind them. Say a description claims a tool runs in Microsoft 365 Copilot, but its setup guide only covers ChatGPT. Trust the setup guide.
Anthropic’s enterprise checklist makes the same point for skills. It tells reviewers to confirm that a skill’s scripts do what the skill says they do.
Check 1
Who made it
- Look for
- A named person or company you could contact
- Warning sign
- No name, or a handle with no other trace
Check 2
A source you can read
- Look for
- The skill files, the setup guide, or the permissions and privacy statement in your admin center
- Warning sign
- A description and nothing behind it
Check 3
Employee data and where it goes
- Look for
- Which data types it handles, which services receive it, and which model runs it
- Warning sign
- External calls or connected services the description never mentions
Check 4
Claims vs documentation
- Look for
- Documentation that supports every capability the description claims
- Warning sign
- "Anonymous," "secure," or "works in Copilot" with nothing in the docs to back it
A hypothetical example
Picture a hypothetical People Ops lead at an 80-person company. She finds a GPT that promises to turn rough manager notes into polished performance reviews. Its description says it’s private.
She runs the four checks. The builder’s profile has no name or contact. The GPT connects to an outside service, and nothing explains who runs that service or what it keeps. Nothing in its documentation backs up the word “private,” either.
Manager notes about named employees would go straight into that tool. She doesn’t need to prove it’s unsafe. It fails three of the four checks, so she keeps looking.
What a library listing does and doesn’t tell you
The AI HR Skills library asks for some of the same things before a tool is listed. A listing needs a specific purpose, a named author, a working source, and language that matches what the tool does. You can read how listings are reviewed in full.
A listing is a starting point, not a guarantee of fit. We don’t automatically test the tools we list, and approval isn’t a legal or compliance review. You still need to check the source, the privacy implications, and your own policies before your team uses a tool. SHRM’s checklist also suggests testing a tool yourself and bringing in IT.
If you’ve built a tool that passes these checks, submit your own skill.
FAQ
What should HR check before using an AI tool?
Four things: a named author, a source you can read, what employee data the tool touches and where it goes, and whether its claims match its own documentation.
Is a skill from a public library safe to use?
Not automatically. A listing tells you a tool met the library’s listing requirements, not that it fits your data or your policies. Anthropic’s documentation says to audit any skill from a source you don’t already trust.
Who should be involved in checking an AI tool?
Whoever owns the employee data, plus IT. SHRM’s checklist suggests bringing in people from IT or legal. It also suggests testing tools in a test environment where you can.
What to do next
Before your team uses a new AI skill, GPT, or agent with employee data, run the four checks: author, source, data, and claims. Write down what you find. A tool that passes all four is worth a closer look. If it fails one, ask its author before it goes any further.
When you’re ready, browse the full library of HR skills and plugins or see other guides in the resources section.
Sources
- Source: Agent Skills overview, Claude Docs
- Source: Skills for enterprise, Claude Docs
- Source: GPTs in ChatGPT, OpenAI Help Center
- Source: Data, Privacy, and Security for Microsoft Copilot, Microsoft Learn
- Source: Plan Data, Privacy, and Security for Microsoft 365 Copilot Extensibility, Microsoft Learn
- Source: AI for HR Tool Evaluation Checklist, SHRM (2026)