How to read this
Skim first. Re-read §1 and §5 before you build anything. The other sections are reference · come back to them when the word shows up in a session and you need a fast definition.
1. Same brain, three doors.
One AI. Three ways in. Most HR leaders only ever use Door 1 · and that's where the ceiling is.
A blank box. You paste, you ask, you get an answer. Nothing carries between conversations.
Claude opens your folder. It reads your files, edits them, and saves them back. It remembers across sessions.
Claude builds something, runs it on a schedule, talks to your other tools. You set the brief and stay in control.
You don't graduate. You pick the door that fits the work. Most of your wins this year will come from moving one recurring task from Door 1 to Door 2 or 3.
2. CLAUDE.md · the file that makes Claude know you.
Every fresh chat, Claude starts from zero. CLAUDE.md is a plain-text file it reads at the start of every session · so it already knows your role, your rules, and your voice before you ask anything.
- Who you are. Role, org size, what you own.
- How you write. Bullets vs paragraphs. Sentence length.
- Frameworks you use. SCALE · CLEAR · IMPACT · READY.
- Vocabulary rules. "We say 'people,' not 'human capital.'"
- Hard limits. "Never include real names or comp figures."
Identity, voice, rules. Stays true across every task.
- One specific procedure (use a runbook).
- Today's data or numbers (paste in chat).
- Step-by-step instructions for one task.
- Anything that changes weekly.
- Anything secret (no API keys, no passwords).
Procedures and data live elsewhere. CLAUDE.md is identity.
Don't try to write the perfect file. Start with three lines. Every time you correct Claude, add the correction. In a month it captures your judgment.
3. Skills · reusable recipes.
A Skill is a small markdown file that teaches Claude how to do one specific job. Install once. Use forever.
A short, kebab-case ID Claude uses to invoke the Skill.
One sentence so Claude knows when this Skill is the right tool.
The instructions Claude follows · format, voice, examples, edge cases.
When to write one: when you've copy-pasted the same prompt five times. That's a Skill in disguise.
4. Routines · AI on a schedule.
A Routine fires automatically. Monday 7am: inbox brief. Friday 4pm: weekly leadership recap. Day 60 of a new hire: nudge to their manager.
You describe what you want in plain English. Claude Code writes the schedule for you. One rule: always have the routine confirm it ran · a single line in your Slack DM is enough. A silent failure is worse than no routine.
5. Runbooks vs CLAUDE.md.
Both are markdown. They do different jobs.
- Loaded: automatically, every session.
- Holds: who you are and how you work.
- Answers: identity, voice, frameworks, hard limits.
- Cadence: updates slowly. Stays stable for months.
"I'm a VP People at a 1,200-person SaaS co. I write in bullets. We use SCALE."
- Loaded: when you point Claude at one.
- Holds: the steps for one specific task.
- Answers: "what do we do when X happens?"
- Cadence: one per recurring task. Grows over time.
"How to run the quarterly engagement survey." Or "How to prep the weekly leadership update."
Rule of thumb: identity goes in CLAUDE.md. Procedures go in a runbook.
6. Why we build with code, not no-code.
Zapier and Power Automate are great for "ping me when X happens." We use code for the recurring, judgment-shaped work · the JD review, the candidate scoring, the exit-interview synthesis.
Three reasons it earns the extra muscle:
- You own it. The file is in your folder. No vendor can change pricing on you or sunset the feature.
- No ceiling. If you can describe it, you can build it. No "the platform doesn't support that."
- Scales for free. Same code runs for 10 employees or 10,000. No per-task billing.
You don't abandon no-code. You add the code muscle so the choice is yours.
7. Git, GitHub, "the repo."
A repo is a folder that tracks every change. A commit is a save with a one-line label. Push sends your commits up to GitHub so others can see. That's 90% of the vocabulary.
You won't be writing pull requests as an HR leader. But your CoS, your Skills, your Routines all live in a repo · and owning your code means owning your repo. We set yours up in Week 3. It's a 5-minute step.
8. The terminal · only the parts you'll touch.
A text window where you type instead of clicking. Two commands get you started:
mkdir my-cos-folder # make a new folder
cd my-cos-folder # step into itFrom there Claude Code handles the rest · and it shows you every step before it runs, so you stay the approver, not the typist. Reviewing Claude Code's plan is the same skill as reviewing a vendor implementation plan.
9. API keys · the password apps use.
When one app needs to talk to another · Claude reading your Granola notes, posting to your Slack · it uses an API key. Treat it like a password. Anyone with the key can act as you.
Keys live in a hidden .env file that never leaves your machine and never gets committed to git. Your IT team handles enterprise keys for anything touching production HRIS data.
10. The four shapes a first project can take.
Pick one. Don't invent a fifth. The first projects that ship in 5 days fit one of these neatly:
Automate
A weekly task that takes 30+ minutes and follows the same shape.
Improve
A process that mostly works but has one painful step.
Build small
A tiny tool your team wishes existed.
Capture tribal knowledge
Turn one person's instinct into a Skill anyone can use.
11. Three questions before you start.
Can you describe it in two sentences? If not, you don't understand it yet. If the second sentence has more than one "and," split the project.
What's the smallest version that still delivers value? Strip everything that isn't strictly necessary. The Inbox Prioritizer didn't start as a CoS · it started as one Skill that summarized one inbox.
What's your first 10-minute step this week? Not this month, not after more reading. Pick a concrete action. Do that one thing.