The rule that mattered most was never in the handbook, and I'd said it out loud for six years

I handle the odds and ends at my company — expenses, badges, onboarding, exits. Every year a few people ask me the same question: how many days of annual leave do I get.

The intern who joined in May is named Zhou. She sits diagonally across from me. That afternoon she came over with her mug and asked how many days she gets this year, since she started in May.

I told her I'd check.

The handbook is on the third level of the shared drive, and it's just called the Handbook. I found the annual leave section. That paragraph is written to be hard: if you've been there a year but less than ten, it's pro-rated against whatever is left of the year, and any fraction of a day is dropped.

I read it out loud. She nodded and said oh. Two seconds later she asked: so can I take it in October.

I couldn't answer that one. I'd never worked out how the pro-rating came out. For the past few years I'd taken the table my predecessor left behind, filled in names and dates, and left the rest alone. The table is at the bottom of a drawer. She left in a hurry the year before last and handed over nothing but that table and one line: just fill it in. The two lines of small print at the top of it are still in her handwriting.

I stayed late that night. Past eight, only my row of lights was still on. I copied the handbook pages out, paragraph by paragraph, and pasted them in. The thing I pasted them into can't reach our shared drive and I didn't expect it to — the only thing I could give it was the text I'd copied.

So I wrote: I handle HR odds and ends at my company. A colleague joined this May, and she had two years of work experience before that. She needs to know two things: how many days of annual leave she gets this year, and whether she can take it in October. Give me something I can send her as a message, three or four sentences, saying clearly how many days she gets, when she can take them, and whether they expire at year end. Don't use words like "pro-rated" or "calendar days". Explain it the way a person would.

The next morning it was in my inbox.

The first version read like a regulation. "In accordance with the rules", "after pro-rating" — both still there, plus a line at the end: if you have questions, please consult the HR department. Which is me.

I changed three things. The dates became "from May to December you'll get about three days". "Pro-rating" was cut. "Expires at year end" became "if you don't use it this year, it's gone".

Then I added a line that wasn't in anything I'd pasted: if you really can't fit it in, it can carry over to March, but don't leave it till then.

That line came from a meeting the year before last. The boss said it once. It never went into any document.

Then I sat and looked at it for a while. In six years I had never once written that rule down. Whoever asked, I said it out loud, and then forgot I'd said it. It isn't in the handbook. It isn't in any document. It's only in my mouth.

I sent the paragraph the next morning. She replied: thanks, that's great.

At the end of February she came by and said she'd take it that month. A few days later the other intern from the same intake asked me. Zhou had forwarded her the paragraph. I looked at it, changed nothing, and forwarded it again. This year someone else asked. I dug the paragraph out, changed two dates, and sent it.

The file and the rule were two different things

The handbook is the rule as published. What actually decided the answer was an override: the exception that lived in speech and never made it into a document. The two are not the same object, and only one of them was findable by anyone who wasn't standing in front of me.

In six years at this company I have never read that handbook from beginning to end. The few lines that actually matter aren't in it. I used to think I remembered them. Every time someone asked I answered straight away, smoothly, so smoothly that I took it for memory.

If you had asked me last year whether our leave policy allowed carry-over into the next year, I would have said yes, immediately, with confidence. If you had asked where that is written, I would have had nothing. A rule with no citation is a rule that lives in a person, and a person is not a document: it has hours, moods, and a two-week holiday of its own.

That is the first thing worth taking out of this, because it is not specific to leave policy. Every team has a handful of behaviours that are load-bearing and unwritten: the exception to the naming convention, the one client whose files go somewhere else, the step everyone skips because it broke once in 2023. They are not in the handbook because the handbook is written for the general case, and these are the cases the general case doesn't cover. They survive on repetition. They get learned by sitting next to someone.

A smooth answer is a reconstruction, not a retrieval

This is the part I keep coming back to.

Six years of answering that question, always quickly, never once checking. I assumed that meant I had it. Writing it down is what showed me that I didn't. What I had was the ability to produce a plausible answer on demand, out loud, and then drop it. Producing it again the next time was not recall. It was rebuilding — from a few remembered fragments, with the gaps filled in by whatever sounded right that day.

The uncomfortable version: I was making it up fresh each time, thinking it through, saying it, and losing it. Accuracy was mostly fine, which is exactly why nobody caught it, including me.

Two things follow that are worth knowing if you work with people, or with systems, or with both.

The first is that fluency is not evidence. Being able to answer without pausing feels like knowledge from the inside, and it feels like knowledge from the outside too — the smooth answer is more convincing than the hedged one. The test that separates them is not "was it right". It is: could you produce it again tomorrow, without the person who asked, without the conversation that prompted it? Did anything survive?

The second is that writing it down is the only way to find out how much of it you were improvising. My paragraph is four sentences. After six years of explaining this to people, four sentences. The parts I was sure about turned out to be the parts I had to go and look up, and the parts I had to look up were mostly the pro-rating arithmetic — the bit that was written down, in the handbook, in the paragraph that is hard to read.

The unwritten part was shorter and simpler than I expected. It just happened to be the part that decided her answer.

What I handed over was my abridgement, and the answer inherited that

I did not hand over the handbook. I handed over three pages I had copied out, which is not the same object.

The thing I pasted them into cannot read our shared drive, or anything else I keep on my machine. It works from what I give it, in words — that is the design, and on that evening it was also the whole story: three pages and one paragraph of instructions.

So the ceiling of the answer was my abridgement. If the page I skipped had said something about March, nobody in that exchange would ever have found out. The output was as complete as the copy, and the copy was as complete as my judgement at eight in the evening about which parts of a policy matter.

There is a second consequence, quieter. A copy carries provenance that the copy does not show. I knew those three pages had come out of the Handbook. The system had no way to know whether the source was authoritative, current, or my summary of a summary of a rule someone half-remembered. Everything it produced sat on top of a source whose status it could not see, and neither could I once the tab was closed.

The one line it could not write was the one line I could not cite. That is not a failure of the tool. That is a map of where our knowledge actually lives.

Written down is a different kind of object

Before that evening, the rule existed as my availability. Ask me and you get it. Don't ask me and it doesn't exist. That is a service with an uptime requirement, and the uptime requirement is one person who also has to take annual leave.

Afterwards it was a paragraph. The second intern asked; I changed nothing and forwarded it. This year someone asked; I changed two dates. The marginal cost went to nearly zero, and — more to the point — the answer stopped depending on whether I was in a hurry that day.

But notice what did not get better. Nobody has ever checked whether the March carry-over is still the deal. It got written down, which makes it look authoritative, which makes it less likely that anyone will ask again. "Expires at year end" sits in a document that gets revised by somebody with a reason to revise it. "Carry over to March" now sits in a paragraph that gets forwarded.

So the paragraph is more useful and less governed than the handbook. That trade is real and I don't have a clean answer to it. The best I have is: when you write down an override, write the date and who said it, next to the rule. Otherwise it will outlive the reason it existed, and the person who inherits it will have no way to tell it apart from policy.

If you delegate work to a system, the exception is the spec

The rule was in the file. The exception was in my mouth. Hand over only the rule and you get an answer that is correct against the document and wrong against the company. Some practical versions of that, which I would have found useful a week ago:

Hand over the exception in the same message as the rule. Not "here is the policy, apply it", but "here is the policy, and here is where we don't follow it". The second half is the half that contains your actual practice, and it is the half that nobody writes down because it feels obvious while you're saying it.

Don't hand over your summary and call it the source. If you abridged something, say what you abridged and what you left out. A summary that presents itself as the whole text removes exactly the information someone would need to notice that it isn't.

Ask for the answer in the shape it will be delivered in. I asked for three or four sentences, suitable to forward, no jargon. What came back the first time was a regulation, because a question about policy pulls toward the register of policy. The format constraint is what moved it, not a request to "be more human" — that request would have produced a regulation wearing a cardigan.

Write down the thing you say most often, not the thing that is hardest to explain. I would have chosen the pro-rating arithmetic, because it is fiddly and I am proud of getting it right. The thing that actually mattered was the sentence I said without thinking, three times a year, in a corridor.

Put the artifact where the next person looks. Mine is currently in a notes file that the next person will never open, which is a better problem than the one I started with and still a problem.

Anything you write down should carry its source and its date. A rule with a citation can be argued with. A rule without one becomes folklore with a timestamp, and folklore is what we had before, only now it's tidier.

Where this sits

CloudPal runs long jobs in the cloud and hands back a finished document. Nothing to install, nothing running on your machine, and nothing read from the files you already keep — it works from what you give it, in words. That is why the three pages were the whole input, and why the one sentence I didn't type stayed mine.

It is in limited preview. New accounts get a dollar of credit, which is enough to run something real. It costs a few times less than the comparable tools, and you pay for what you use.

In six years I never wrote down the rule I gave out most often. It took an intern with a mug and a machine that can only read what I type for me to notice. The paragraph is four sentences long. The version in my head was never that short — it was just spread across six years and a few hundred conversations, and every one of them ended with me going back to work.

CloudPal.TuringCorp.net — create an Agent Pass, sign in, and say what you want done.

Earlier in this series: half an instruction, four blank fields, and three weeks nobody checked, it guessed a number, we told it off, and every number after that got checked, we never wrote our format down, so it took our old drafts as the spec, correct is a low bar, the same three facts, in a different order, the second half of the draft was policy we never agreed to, it can't read anything on my laptop, a direction is not a spec, four sentences in, seventeen pages out, the requester is not the reader, the first run is a demo, the ninth is the product, the meter, not the price, we deleted the install step.