AI agents for HR: the skills, the job and the agent
Most HR teams use AI one text at a time: a job ad here, a tricky email there. Everyone asks in their own way, and nobody has agreed what the AI may see about the people involved.
There is a better way, and it is a ladder with three rungs. A skill writes down one task the way your best HR partner does it. Several skills in order make a job, with someone checking each step. The job running on its own, stopping for a person before anything leaves the house, is an agent. Below is that ladder for HR and recruiting, with the rules for personal data built in.
- 01
6 skills
One task each, described in detail.
- 02
Connectors
Reach into the systems the tasks need.
- 03
One job
Onboarding a new starter
- 04
One agent
The onboarding agent
Claude skills for hr: the 6 to write first
job-adJob ad
Use when: a role is approved and the job ad needs writing or updating.
Takes the role, team, must-haves and nice-to-haves, writes the ad in the house template, and cuts requirements the role does not need and words that signal age or gender (“young”, “digital native”).
Output: A job ad for the hiring manager to approve, with the must-haves listed on their own.
interview-kitInterview kit
Use when: interviews for a role are booked and the panel needs its questions.
Builds a structured interview from the job ad: the same questions for every candidate, what a strong answer looks like, and a scoring scale for each requirement.
Output: An interview guide and a scorecard the whole panel uses.
candidate-replyCandidate replies
Use when: a candidate needs an update, an interview invitation or a no after a decision a person has made.
Drafts the message from the decision and the stage, warm and clear, and gives no reasons the hiring team has not agreed to give.
Output: A draft for the recruiter to read and send.
onboarding-planOnboarding plan
Use when: someone has signed and has a start date.
Turns the onboarding checklist into a plan for this person: tasks for HR, the manager, IT and the buddy, each with an owner and a date, from before day one to day 30.
Output: A dated plan with the first week laid out day by day.
welcome-emailWelcome email
Use when: an onboarding plan is ready and the new starter should hear from you before day one.
Writes the welcome email from the plan: time and place on day one, who meets them, what to bring and what the first week looks like.
Output: A draft welcome email, ready for HR to send.
policy-answerHandbook answers
Use when: an employee asks about leave, holidays, expenses, benefits or another rule in the staff handbook.
Finds the answer in the handbook and policies, quotes the section it comes from, and says so plainly when the handbook does not cover the case.
Output: A short answer with a link to the source, or a note that HR needs to answer.
What a good skill looks like
An example written for this page, following Anthropic’s guidance for writing skills. Adapt it to how your team works.
---
name: onboarding-plan
description: Turns the onboarding checklist into a dated plan for one new starter, up to day 30. Use when someone has signed and has a start date, or a manager asks what must happen before they start.
---
# Onboarding plan
## What you deliver
One plan, four sections, in this order: Before day one, Week one (day by day), Days 8 to 30, Open questions.
Every task has an owner (HR, manager, IT or buddy) and a date.
## Steps
1. Read the new starter form: name, role, team, manager, start date, place of work, equipment.
2. Read the onboarding checklist and the team's own pages. Do not add tasks that are not in them.
3. Set every date from the start date. Count working days and skip public holidays.
4. Before day one: contract signed, accounts and equipment ready, buddy named, welcome email sent.
5. Week one: one line per day, with each meeting and who leads it. Keep day one light.
6. Days 8 to 30: what the checklist sets for the first month, such as the first one-to-one and the day-30 check-in.
7. List everything the form did not answer under Open questions, with who should answer it.
## Rules
- Never guess a date, a manager or a buddy. If one is missing, write "Not set" and add it to Open questions.
- Personal data: use only name, role, team, manager, start date and place of work. Never ask for or write down personal ID numbers, bank details, health information, allergies, union membership or family circumstances. If the form contains any of them, leave them out and tell HR.
- If the new starter has asked for adjustments, write only "Adjustments: HR to agree with the new starter".
- The plan goes to HR, not to the new starter. HR decides what is shared.
## Example
New account manager, Sales, Acme AB, starts 9 June.
- 4 June, IT: laptop and accounts ready.
- 9 June: 09:00 welcome with HR, 10:30 coffee with the buddy, 13:00 CRM walkthrough with the manager.
- 9 July, manager: day-30 check-in.
- Open questions: buddy not named yet (manager to answer by 2 June).
- The description names the moments
- Claude reads only the name and description until a task matches. This one names the two moments HR actually meets: a signed contract with a start date, and a manager asking what has to happen first.
- A list of what to leave out
- The rules name the personal data the skill must never write down. “Be careful with personal data” changes nothing; a list does.
- Gaps become open questions
- A missing buddy or start time goes under Open questions with a name next to it. The plan never fills a gap with a guess, so HR sees exactly what to chase.
From skills to a job: Onboarding a new starter
Today HR rebuilds the same plan for every new starter from an old checklist, and something is always forgotten: the laptop, the buddy, the first one-to-one. With the skills, one HR partner has the plan, the welcome email and the first week ready in under an hour, in Claude or ChatGPT.
- 01Give the AI the new starter form and the onboarding checklist.
- 02Run onboarding-plan to get tasks, owners and dates up to day 30.
- 03Run welcome-email on the plan.
- 04Read both and add what only you know, such as who the buddy is and what the team is in the middle of.
- 05Send the email yourself, book the first week’s meetings and give the manager their tasks.
From job to agent: The onboarding agent
A hiring manager fills in the new starter form in Typeform.
- 01Reads new responses to the new starter form: name, role, start date, manager, team and equipment. via Typeform MCP
- 02Reads the onboarding checklist and the team’s pages in the staff handbook. via Notion MCP
- 03Runs onboarding-plan, then welcome-email on the plan.
- 04Creates the new starter’s onboarding page under the page HR named. via Notion MCP
- 05Creates the welcome email as a draft in the HR partner’s Drafts. The agent never sends it. via Gmail MCP
- 06Posts in the HR channel that the plan and the draft are ready for review, with the start date and manager. via Slack MCP
The HR partner reads the plan and the welcome email, fixes what is wrong, sends the email and shares the page with the manager. Nothing reaches the new starter before that.
A finished onboarding plan, a welcome email in Drafts and a clear list of who does what, within a working day of the form coming in.
In Claude, make it a scheduled task: describe the job once, choose the cadence (every morning works here), and it runs with your skills and connectors even when your laptop is closed. It checks for new forms on that cadence; it does not react the moment a form is sent. Scheduled tasks are on every paid Claude plan. In ChatGPT Work, a scheduled task does the same on the same cadence.
With Deployed OS, HR’s skills live in one place and an update reaches everyone the same day, in Claude, ChatGPT and Copilot. Each person connects with their own login, so the AI sees only what that person may see. Admins decide which groups may let the AI write in Notion, Gmail or Slack, every call is logged, and you choose where the data is stored, down to the city.
Where it goes wrong
- Letting the AI decide on candidates
- Do not let an agent reject or rank people on its own. A person makes every decision about a candidate; the AI drafts the interview kit before it and the reply after it.
- Personal data in the skill
- A skill is shared and reused. Keep names, salaries, health information and notes on individuals out of it: the rules belong in the skill, the person’s details in the task.
- A form that asks too much
- If the new starter form asks for personal ID numbers, bank details or allergies, every agent that reads it sees them too. Ask only for what onboarding needs and keep the rest in the HR system.
- Answers without a source
- A policy answer without the section it rests on is a guess with HR’s name on it. Make the skill quote its source and hand over to HR when the handbook is silent.
AI agents for hr: common questions
The recurring work with fixed steps: job ads, interview kits, onboarding plans, welcome emails, answers from the staff handbook. It prepares the material in the same order every time. A person makes every decision about people and sends everything that goes out.
Start with onboarding-plan or interview-kit: you use them for every hire, and they hold little personal data. Use the skill by hand on the next few hires, fix what it gets wrong, and build the agent only when you trust the result.
It can prepare them, not send them. In Gmail and Outlook it creates drafts that are never sent. In Google Calendar it can create events without sending invitations, and in Calendly it creates a single-use booking link that a person shares. In Notion it can create a page, and in Slack or Teams it can post a message. Typeform, Confluence, Google Drive and SharePoint are read-only. We recommend that a person sends every message to a candidate or new starter.
Either works. A skill is plain text, and through Deployed OS the same skills and connectors work in Claude, ChatGPT and Copilot. Pick the assistant HR already uses.
Keep three rules. Put no health data, ID numbers or salaries in a skill, and do not let the AI read them unless the task needs it. Let no agent make a decision about a person on its own, such as rejecting a candidate. Ask only for the data the task needs. With Deployed OS each person connects with their own login, so the AI sees only what they may see, every call is logged, and you choose where the data is stored, down to the city. Deployed is an EU company and fully GDPR compliant. This is not legal advice: check your setup with your data protection officer.
It depends on the size of your team. Get in touch and we will walk you through it.
Start with one job.
Tell us which task eats your team’s week. We’ll show you the skills that do it, and the agent that runs it.
