Wednesday, 4pm. Slack: "Quick one, if I work from Portugal for a month, does anything change with my days off?"
You know the answer. Roughly. It's in a 2023 email thread, a Notion page still titled DRAFT, and your head, and the three don't quite agree. So you write the reply. Eleven minutes. Third time this quarter.
That's the cost of not having a handbook. You are the document, and every question is an interrupt.
You've probably tried the obvious fix already: open a chat, type "write an employee handbook for a 60-person software company." What comes back is competent and useless. Fifteen sections that fit every company and describe none. A PTO policy borrowed from nowhere. "Ask HR" in every third paragraph. A template with the blanks pre-filled, which is worse than a template with blanks, because now the wrong numbers look decided.
Generic in, generic out. The handbook is only useful if it says how your company works, and the model doesn't know that unless something makes it find out.
That's what this lesson adds: a prompt that forces the research before the questions, the questions before the structure, and the structure before a word of policy. Plus the context you built in Module 1, so it starts from your rules instead of from zero. The result is a first usable handbook in one sitting, one a person joining next Monday could actually run on.
What you'll have at the end
One document, saved where your team reads. Its shape depends on your size, and Claude proposes the shape, so know the three before it does.
Under about 30 people. Ten to fifteen pages that read like a letter from the founders. How we work, time off, pay, growth, expenses, conduct, leaving.
30 to 300 people. Twenty to forty pages. Shared rules plus a short appendix per country for what legally differs: statutory leave, holidays, notice.
300+, or regulated. Everything above plus compliance, grievance, whistleblowing, conflicts of interest. At this size the handbook is a plain-language map over formal policies. Summarise and link, don't duplicate.
Whichever tier, every section answers three questions: what the rule is, how it works in practice (where to request, who approves, which tool), and who to ask. Every fact traces to your website, an existing document, or your answer. Where it's none of those, the text says [TO CONFIRM: what needs deciding] and moves on. A made-up notice period is worse than a blank, because a blank gets fixed and a made-up number gets enforced.
All the [TO CONFIRM] lines collect in an "Open items" list at the end. That's your homework, and it's short.
And the sections that depend on employment law are marked, with one plain sentence: these need review by someone qualified in each country you employ people in. Claude isn't that person. Neither is this course.
Before you start
Three things from Module 1 need to exist:
→ The workspace, from How to organize context files. HR-CONTEXT-FOR-AI/ with company/ inside it.
→ The AI policy, from Basic setup, identity and your AI policy. ai-usage-policy.md at the top of the workspace. The handbook touches pay, leave and conduct; the policy keeps names and personal data out of the build.
→ At least one connector, from the same lesson. Email, and ideally where your documents live: Drive, SharePoint, Notion, Confluence.
The process works with nothing connected: web research, interview, output as a file. But every policy Claude can find is one you don't have to explain. Connect what you can, then start.
How the build runs
Three steps, and each ends with a pause where you confirm.
The pauses are the design. The build stops after research, so you can correct what Claude thinks it knows about your company, and after the proposed structure, so you approve the table of contents before any policy gets written. Wrong assumptions caught here stay cheap.
Step 1, research before questions. Claude reads your website and searches the web for headcount, locations, culture signals. Then, with your okay, it searches your connected tools for what already exists: policies, code of conduct, offer letters, an old handbook nobody finished. It reports back in plain prose.
Research first changes the interview. "How many people do you have?" is a question for a stranger. "Your careers page says remote-first across the UK and Poland, still right?" respects your time. And anything already decided is carried in as-is. Your 2023 PTO policy is the PTO policy. Claude confirms, it doesn't reinvent.
Step 2, interview, then structure. Up to ten questions, fewer if research answered them, all in one message. They're picked by gap from a bank of about twenty, and they go after what no website has: how time off actually gets approved, what happens when someone underperforms, how pay decisions get made, and what the company deliberately does not want to be. That last one shapes tone more than a values page does.
Answer what you can. Half-answered is fine; the rest becomes open items. Then Claude proposes a table of contents, one line per section with where the content comes from. You say yes or edit. Nothing gets written until you do.
Step 3, write and save. The whole handbook in one pass, in your company's voice as far as the site and documents show it. Then it saves where your team will find it: wiki, file storage, or a downloadable file. Before creating anything in a shared workspace it states the location and waits. A new page in the company Notion is visible to everyone, and you decide when it appears.
The prompt
Start a Cowork task with your HR-CONTEXT-FOR-AI folder connected, in the Handbook & Policy room from How to organize context files. Every later update to the handbook starts from this history.
The prompt lives in Effy's HR skills library on GitHub, so the version you copy is the current one:
https://github.com/Effy-AI/hr-skills/blob/main/employee-handbook.md
Copy the whole file, paste it in, add your domain on the last line.
Forty-five minutes to an hour, most of it answering questions about your own company.
What to watch for
The research report. Read it slowly. This is where "we're a 60-person company" becomes "60 people plus 12 contractors in Brazil the handbook needs to cover or exclude." Every policy downstream inherits what you let through here.
The existing-handbook switch. If Claude finds one, even half-finished, the job becomes updating. That's right. A rewrite throws away decisions someone made for reasons you may not remember. The original stays untouched.
The table of contents. Your one cheap edit point. Renaming a section costs nothing here and a lot after twenty pages. If a section exists only because the template had it, cut it.
The plausible default. The most helpful-looking thing a model can do with a gap is fill it with a number that sounds right. The prompt forbids this. Check anyway: any number you didn't give and no document contains is invented.
Other people's policies. The public handbooks in the prompt are for structure and tone, not values. If a section reads like a well-known remote company's, ask where it came from.
Legal sections. Time off, pay, conduct, leaving, every country appendix. Reviewed by someone qualified in each country before it goes live, every time. A handbook is a promise to employees, and promises in the wrong jurisdiction are expensive.
Where it lives
The handbook lives where people read it. Not in HR-CONTEXT-FOR-AI/.
What goes in HR-CONTEXT-FOR-AI/ is one line in company/policies-index.md: name, location, owner, last updated. Copies drift, and a stale copy of a handbook is wrong with confidence about someone's leave. With the pointer in place, every future task reads the real thing: the HR Helpdesk room answers from it, the onboarding plan links to it, the offer letter matches it.
Keeping it true
The handbook that gets abandoned is the one nobody owns. Question 22 in the interview exists for that. Answer it.
Then two habits from How to organize context files. The weekly context review from that lesson already reads logs/feedback-log.md and your context files; add the handbook's location and it flags drift, like a new country in your email with no appendix. And every question the handbook should have answered goes in logs/feedback-log.md. Ten of those and you know which section to write next.
That Portugal question? Next time it has an answer before it gets asked.