The mental model for working with Claude: where Claude lives, the four building blocks, what goes where, and the rules for employee data.
Where Claude lives: Chat, Cowork, and Code
Claude shows up in three modes, on web, desktop, and mobile.
Chat. You type, it answers. Quick drafts, questions, thinking out loud. The ChatGPT-like experience.
Cowork — the one that matters for automating HR. Claude doing a real task for minutes or hours while you do something else. It reads your files, plans the steps, writes the output (docs, spreadsheets, decks), and asks you questions as it goes.
One thing worth knowing: sessions keep running in the background. And since July 2026, Cowork tasks run in the cloud — start one, close your laptop, it carries on and pings you when it needs a decision.
Code. Claude in your terminal, mostly for developers. It's the raw version: originally a terminal tool, it works directly in any folder, with fewer guardrails and more control over how the system is wired together. Day to day, a non-coder interacts with it the same way — plain language either way. "Go through the 200 offer letters in this folder, pull name, salary, level, and start date into one spreadsheet, and flag any letter that's missing a signature." If you don't code, skip it — Cowork does the same job with more guardrails.
The quick test: want a thinking partner? Open a Chat — a back-and-forth you steer, you're in it the whole time. "Help me word this tricky message to a manager."
Want a finished thing, and the work touches your files or tools? Start a Cowork task — a job you hand off: describe the goal, point Claude at your files, say what "done" looks like, and it comes back with a finished result. "Read the review exports in this folder and build a calibration summary by team."
In short: Chat helps you write an email. Cowork runs your whole review-comms sequence, builds the handbook, and drafts fifteen development plans while you're in a 1-1.
The four building blocks
Connectors — how Claude reaches your tools
They're the link that lets Claude read from and write to your tools, so everything below works with real data instead of waiting for you to paste it in. For HR the ones that matter: Gmail, Outlook and Calendar (the endless employee and candidate email, 1-1s, review-cycle dates), Drive or SharePoint (where the handbook and pay sheets live), Slack or Teams (employee questions and announcements), Notion or Confluence (your internal wiki), and your HRIS/ATS (headcount, roles, hiring pipeline).
Skills — a shortcut you call
A skill is a procedure: a step-by-step way of doing one task, written down once and ready in every conversation. Not knowledge — method. "How we turn a finished review into a 90-day plan: 2–3 focus areas, 2 actions each, one measurable, check-in dates" is a skill; the review itself is not. You trigger one by name (/performance-review, /pdp, /1-1-prep, /handbook-update, /offer-comms) or Claude picks it up on its own when it spots the task. Because the procedure is written down, the task comes out the same way every time — that's the point.
Plugins — building blocks you install instead of build
Not a sixth block — a shortcut to the five. A plugin is an installable bundle of skills and connectors, packaged by someone else, added in one click. Anthropic's built-in Human Resources plugin is the HR example: one install gives you nine ready skills (/performance-review, /comp-analysis, /draft-offer, /onboarding, /policy-lookup…) and hooks up the main connectors (Gmail, Calendar, Slack, Notion). Start with a plugin, feel the value, then add your own skills for what's missing.
A note on plans: if your organization uses Claude on a Team or Enterprise plan, your admin manages which connectors and plugins are available to you. Don't see the HR plugin or a connector you need? That's a request to your admin, not a missing feature.
One line to remember: connectors bring the data, projects hold the room, context docs hold the knowledge, skills hold the method, scheduled tasks run it on a clock — and plugins let you install the blocks instead of building them.
Scheduled tasks — work that runs itself
A job you set up once that Claude runs on a schedule and delivers finished, in the cloud, even with your laptop closed — a Monday "people brief," a Friday list of managers who still owe reviews.
Keep the split clean: the task holds the when and what — schedule, scope, where to deliver ("every Friday at 4, check review status, leave a draft nudge for me to send"). The method — how to check, how to word the nudge — belongs in a skill the task calls. If the how lives inside the task prompt, you can't reuse it anywhere else and you'll maintain two versions of the same procedure.
Projects — a place you go to
A saved workspace that remembers a set of documents, a set of rules, and its own notes. Everything you do inside it already has that context. Think of it as a room set up for one job you do again and again — a "Review Cycle" room that already knows your form, your competency framework, and your comms rules; a "Hiring" room with your JD templates and scorecards; an "HR Helpdesk" room that answers only from your policies.
A project holds two things: the documents Claude references, and the instructions for that room — rules like "take out names," "never show individual scores in a group summary," "flag anything a lawyer should check."
Context docs — what Claude knows about you
Plain files that hold your standing knowledge, so you stop re-explaining yourself. Some live above all projects (an ABOUT-ME/ folder: about-me, my-rules, my-templates), some inside a project room — the handbook, the pay bands, the review form, the competency framework, the org chart, past survey results. A my-rules.md that says "take out names; anything touching pay or legal gets flagged for human review" is obeyed in every task, without being asked.
The shared-doc problem, in short: if the same document keeps showing up in several rooms, don't upload copies — copies drift, and a stale copy is wrong with confidence. Keep one source of truth (in a connector, or in a folder above the projects) and point every room at it. More on this in section 4.
.svg){kind=icon}