How to build Claude skills for people ops workflows

Most people ops teams that start using Claude get stuck at the same point: ad-hoc prompting works, but the useful prompts live in someone’s notes app, every new hire gets a slightly different checklist, and the output depends on who typed the request. According to AIHR, 57% of HR professionals’ time goes to administrative tasks. That is repeatable work, and repeatable work is what skills are for.

A Claude skill is a saved set of instructions that tells Claude exactly how to run one people ops task, step by step, every time. Connect it to your HR system through MCP (Model Context Protocol, the standard that lets AI tools read and write your other systems) and the skill works on live employee data instead of whatever you remembered to paste in. This is Vol 2 of our Claude for people ops series: Vol 1 covered what Claude can do and how to set it up; this one is hands-on, with five skills you can copy, paste and use today against your own employee data and Slack.

TL;DR

• A skill turns a good people ops prompt into a repeatable workflow: same steps, same checks, same output format every time
• Every good people ops skill has five parts: a trigger, the data it needs, the steps, the checks, and where a human approves
• Skills get far more useful when Claude can read and update your HR system via MCP, so the data is live and the actions are logged
• Below are five copy-paste skills: new hire setup, probation check-in tracker, leave balance audit, monthly pre-payroll report, and a review-cycle readiness check
• Start with the one task you repeat most, run it manually for two weeks, then tighten it before adding the next

What is a Claude skill, and how is it different from a prompt?

A prompt is a one-off request. A skill is a reusable instruction file that Claude loads when the task comes up, either because you call it by name (for example /new-hire-setup) or because Claude recognises that your request matches the skill’s description.

The difference matters in people ops because consistency is the point. If two managers onboard two new hires, the checklist should be the same. If you audit leave balances every quarter, the audit should check the same things each time. A skill holds that standard so you don’t have to.

Skills are written in plain language, so you don’t need to be technical to write one. Paste any of the prompts below into a Claude chat and ask Claude to save it as a skill; its built-in skill creator handles the file format. You can also save it yourself as a SKILL.md file if your team keeps skills in a shared folder.

What makes a people ops skill work well?

Every skill in this post follows the same structure. If you write your own, use these five parts:

• Trigger: when should this skill run? (“When a new hire’s contract is signed”)
• Inputs: what data does it need, and where does it come from? (the contract, the employee record, the manager’s name)
• Steps: what does Claude do, in what order?
• Checks: what should Claude verify or flag before finishing? (missing manager, negative leave balance, start date in the past)
• Approval point: where does Claude stop and ask a human before acting?

The approval point is the part people skip, and it is the part that makes a skill safe to use with employee data. A good rule: Claude can read anything, draft anything, and update low-risk fields on its own. Anything involving pay, contract terms, terminations or messages to employees waits for a human yes.

Why connect your skills to your HR system via MCP?

A skill without data access can only work on what you paste in. That means exporting a CSV, cleaning it, pasting it, and hoping it is still current. It also means Claude’s updates go nowhere: you copy the output back into the system by hand.

With an MCP connection, Claude reads and updates your employee records directly. Taito.ai’s MCP server, for example, lets Claude list employees, read time off balances, create records, assign policies and managers, and pull payroll reports, all with the same field-level permissions you have in the app. Every change Claude makes appears in the activity log, attributed to you.

One of our customers, a roughly 30-person energy company, set up their whole company this way: employees, positions, managers, locations, and time off and attendance policies, through Claude, in a single evening.

Where should your skills get their employee data?

Every skill needs one source of truth: a single place where employee records, managers, contracts and time off balances are kept up to date. That source is the context Claude works from. If it is out of date or split across tools, the skill’s output will be too.

The prompts below use Taito.ai as that source of truth, because it holds employee data, time off, attendance, documents and review cycles in one place and connects to Claude through MCP. But the skills don’t depend on Taito.ai. You can point them at wherever your employee data lives today:

• Another HR system with an MCP server: the steps and checks stay the same; only the action names change
• A Google Sheet or Excel file: connect it to Claude or upload it, and the read-only skills (probation tracker, leave audit, payroll summary) work as written
• A CSV export from your current system: fine for a one-off check, but you will be working on a snapshot, not live data

To switch, change the Taito.ai (via MCP) line in each prompt to your own source, for example “from the employee sheet I’ve shared”. The skills that create or update records (new hire setup) need a system Claude can write to, so with a spreadsheet they will draft the changes for you to apply by hand.

How do you automate new hire setup with Claude?

Use it when: an offer is signed and the new hire needs to exist in your systems before day one.

What it does: reads the signed contract, creates the employee with the right manager and time off policy, and gives the manager a day-one checklist, so every new hire is set up the same way without retyping contract details by hand.

---
name: new-hire-setup
description: Set up a new hire from a signed contract. Use when someone shares a signed contract or says a new person has signed.
---

Set up a new employee in Taito.ai (via MCP) from the signed contract I share.

1. Pull out their name, email, job title, team, manager, start date, country and probation length.
2. Show me a summary of what you'll create and wait for my "go".
3. Create the employee, assign their manager, and assign the time off policy for their country.
4. Attach the contract to their documents.

If you can't find the manager or team, ask me. Don't send the employee invite until I say so.

Finish with a short day-one checklist for the manager.

How do you track probation check-ins with Claude?

Use it when: you want a weekly check of who is coming up to their probation review.

What it does: lists everyone whose probation ends in the next 30 days and drafts a reminder for each of their managers. A missed probation date is one of the most common, and most avoidable, people ops mistakes in a growing company, and this skill catches it a month ahead.

---
name: probation-tracker
description: Find employees whose probation ends in the next 30 days. Use for the weekly probation check.
---

Using Taito.ai (via MCP):

1. List everyone whose probation ends in the next 30 days, with their manager and end date.
2. Flag anyone with no manager, or whose probation has already ended without a review.
3. Draft a short Slack message to each manager with the employee's name and the date.

Show me the list and the drafts. Don't send anything.

If you don’t have a probation review format yet, our lightweight probation review template works well alongside this skill.

How do you audit leave balances with Claude?

Use it when: before the end of a holiday year, or quarterly. In Finland the holiday credit year runs from 1 April to 31 March, so a February audit catches unused leave before it becomes a carry-over or pay-out problem.

What it does: checks every employee’s time off balance and flags anything unusual: negative balances, people sitting on too much unused leave, and anyone missing a policy, so you catch leave problems while there is still time to fix them.

---
name: leave-balance-audit
description: Check time off balances across the company and flag anything unusual. Use before the holiday year ends or quarterly.
---

Using Taito.ai (via MCP):

1. List every active employee's time off balances.
2. Flag negative balances, balances over [X] days, and anyone with no time off policy.
3. List who should be nudged to book leave, highest balance first.

This is read-only: don't change any balances. If something looks wrong under local leave law, flag it and explain why instead of fixing it.

The leave calculation itself should come from the system of record, not from Claude. Taito.ai calculates Nordic leave rules natively, so the skill’s job is to spot anomalies, not to redo the maths.

How do you prepare monthly payroll data with Claude?

Use it when: the week before payroll closes, so your accountant or payroll provider gets clean data on the first try.

What it does: pulls last month’s payroll data, lists what changed (starters, leavers, salary changes, leave), and flags gaps before anything goes to the accountant, so you stop the back-and-forth emails that follow when payroll data arrives incomplete.

---
name: pre-payroll-report
description: Prepare the monthly payroll summary for our accountant. Use at month end.
---

Using Taito.ai (via MCP), for last month:

1. Pull the payroll report.
2. List what changed: starters, leavers, salary changes, unpaid leave and sick leave.
3. Flag anyone missing a salary, and any time off requests still waiting for approval.

Give me a short summary for the accountant, the flags, and the data as a table. Don't send anything or edit any salaries.

How do you check a performance review cycle is ready to launch?

Use it when: a week before launching a review cycle.

What it does: checks who is in the cycle, flags setup problems, reviews the questions, and drafts the launch message. Most review cycles stall on setup, not on the reviews themselves: missing managers, people who joined last week, or forms nobody has checked. This skill finds those before your team does.

---
name: review-cycle-readiness
description: Check a performance review cycle is ready to launch. Use the week before launch.
---

Using Taito.ai (via MCP), for the review cycle I name:

1. Check everyone in the cycle has a manager.
2. Flag anyone who joined in the last 60 days or is on long leave.
3. Flag any review questions that are vague or overlap.
4. Draft a short Slack launch message with the deadlines.

Show me a pass/flag checklist and the draft message. Don't publish the cycle or send anything.

How do you roll skills out without creating chaos?

Pick the one task you repeat most often and build that skill first. For most small teams that is new hire setup or the pre-payroll report. Then work through it in order:

1. Run it manually for two weeks. Call the skill yourself and read every output. You will find the edge cases quickly: contractors, part-time staff, people in a second country.
2. Tighten the checks. Each time the skill gets something wrong, add a rule. Skills improve fastest from real failures.
3. Share it with your team. Once it is reliable, share the skill so a manager or a co-founder can run it too. The output stays consistent no matter who runs it.
4. Only then add the next one. Five half-working skills are worse than one you trust.

The failure mode to avoid is the opposite order: writing five skills in an afternoon, trusting none of them, and going back to doing the work by hand. Once a skill runs reliably and you are always clicking “yes” at the approval step, it becomes a candidate for running on its own, on a schedule or when something changes in your data. That is a later step, not a first one.

Skills work because they move people ops knowledge out of people’s heads and into a written, versioned process, with a human still approving anything that matters. Taito.ai’s MCP server connects Claude to your employee records, time off, attendance, documents and review cycles, with the same field-level permissions as the app and every change logged. All five skills above run against your real data, not a stale export. See how it works, or we can walk you through it in 30 minutes.