On a Friday night in Japan, mid-vacation, my phone buzzed with a text from one of our sales reps. A prospect she'd been working with for months was suddenly ready to move. They wanted a proposal and a draft contract by Monday morning.

I sat down at a Starbucks in Kyoto and opened my laptop. The plan was modest: write one document under pressure, get back to my family, salvage the rest of the trip.

I walked away three days later with seven things.

The contract she'd asked for. A service definition document we'd never put in writing in twenty-five years of running the company. A reusable proposal skill that takes a discovery transcript and produces full HTML proposal blocks — pain points, pricing narrative, what's included, what changes when you work with us. A pricing formula that back-solves line items from a stated seat price. A second contract format for a different kind of engagement. An internal team reference. And — this is the one I want you to remember — a single-page sales reference for the rep herself, written in business language with no tool names, that she could put in front of a prospect tomorrow.

The one-page sales reference is the whole point of this story. It does not exist in the universe where I just wrote the contract. Nobody would have asked for it. The rep didn't know to want it. But once the system was running — once Claude had absorbed the service definitions, the pricing logic, and the contract terms — producing the sales reference was three more sentences of prompting.

A disclosure, before the math

Before I tell you what made the cascade possible, I owe you a disclosure.

That Friday night in Kyoto was not the work. The Friday night was the withdrawal on work I had done earlier. For about three months, most days, I had been showing up at 4 or 5 a.m. before my family woke up, opening Claude, and talking to it. Not building anything in particular. Not asking it to do a task. Just talking. Sometimes about a real problem. Sometimes about a frustration. Sometimes about a thing that wasn't even work — the bonsai trees, the book, the dogs.

I was overdoing it. Heavy reps. Probably more than I needed to in order to learn the tool. I want to be honest about that, because if I leave it out, this piece reads like every other breathless AI productivity story — guy sits down, magic happens, look how easy. That's not what happened.

But here's the part that matters more than how much I did. You don't need to do what I did. The amount of compounding available to you for one hour a day with the door closed is the same compounding I got with more. I'm not the proof of concept. The hour is. I just happened to take a lot of them.

The boring math

Compounding gets talked about a lot in finance and almost never in knowledge work, and the reason is structural. In finance, compounding works because the previous period's returns stay in the account. You don't have to re-earn last quarter's gains. They sit there, doing nothing visible, generating the next gain.

In most professional work, the previous period's returns don't stay in the account. They get consumed.

You solve a problem on Tuesday. Wednesday you solve a different problem. The skill you built on Tuesday is in your head; it's not retained anywhere. The next person who has Tuesday's problem doesn't have your skill. Your firm doesn't compound — you compound, while you're awake, until you leave. Then the curve resets.

I'd been on the wrong side of that math for twenty-five years. Hard work, good work, but consumptive. Every fire I put out was a fire I'd have to put out again with a different name on it next quarter. The skills I'd built were real, but the artifacts — the documents and processes and reusable structures that should have accumulated under twenty-five years of work — were thin. The conversations I'd had to produce them were gone the day they ended.

The first time I used Claude for a serious problem, this changed. Not because Claude is magic. Because what Claude leaves behind is retained. The conversation becomes a skill. The skill becomes a workflow. The workflow runs again next time without you. The substrate accumulates.

The math finally had a balance to compound on.

Which brings us back to that Starbucks. Three months of substrate, sitting in a project file, waiting. A sales rep with a text and a Monday deadline. The cascade that started with "write the contract" and ended with "and here's a sales reference your rep can put in front of a prospect tomorrow" wasn't impressive. It was inevitable. By the third day there was nothing left to ask for that wasn't a free side-artifact of work I'd already done.

The deck I built about all of this — the one I gave to my team last week — was itself an instance of the practice it describes. I came to the session with a frustration. I walked out with the deck.

Two receipts

Two examples of what the substrate actually looks like, since the abstract version of compounding never convinced anybody of anything.

The first was a one-time contractor doing specialized work for one of our clients. About fifty time entries a month. The "right" path was to train her on the time-tracking system we use — videos, a one-on-one walkthrough, an hour a day of her perfectionism, then me cleaning up her mistakes before invoices went out. I'd been planning to do this for weeks. The thought of it was exhausting.

Instead I built a stateful skill. She keeps her timesheet however she wants. Claude reconciles it, deduplicates, posts the entries to our system via API, writes the status back to her sheet. Picks up the next run where it left off the last one. The skill is still running. I haven't touched it since I built it.

This isn't a time-saved story. It's a bullshit-removed story. Not every compounding win shows up in a dollar number. Some show up as tasks that no longer exist.

The second was a foundational system we'd been meaning to migrate off for five or six years. Not because the migration was complicated. Because I was complicated — the expertise lived in my head, the system worked well enough, and the cost of stopping to migrate always felt higher than the cost of continuing to limp. Single point of failure: me.

One afternoon: export scripts, sanity check, complete catalog of every record, recommendation for the new platform, account walkthrough on the new platform, PowerShell scripts to do the actual migration via API. I'd planned to hand the execution to a junior tech. After two test runs, I ran it myself.

What I sent the junior tech to finish the job was a Claude-drafted email with step-by-step instructions for the registrar updates. The lesson buried in that email is the same lesson as the one-page sales reference back in Kyoto: when you've already done the work to understand something with Claude, the writeup is two more sentences of prompting. Don't explain it twice. The handoff is free.

Two receipts. Both rhyme. Both end on a free side-artifact that nobody asked for and that wouldn't exist if the substrate weren't already there.

The one condition

I told you at the top there was one condition the math requires. Here it is.

An hour. With the door closed. Outside the office, away from the stimulus, away from the pings. Talking to the tool the way you'd talk to a friend you call on your drive home — to complain about your day.

That's the whole prescription.

It does not work if you sit down at your desk between meetings, with one ear on Slack and one eye on the inbox, and decide to "try Claude for ten minutes." That's not deep work. That's auditioning for deep work in front of an audience that is constantly interrupting the audition.

The reason the door has to be closed isn't that you need silence. It's that you can't be honest in front of the room. And the honesty is the part that produces the help.

Most professional work has trained you, for twenty years, to send finished thoughts. The polished question. The cleaned-up summary. The version that doesn't show its work. Claude can respond to a finished thought, but the response will only be as useful as the thought you sent. If you send the conclusion, you get the conclusion back, slightly polished. That's the experience most people had when they tried these tools once and bounced — they were sending finished thoughts and getting finished thoughts back, and they correctly concluded that wasn't worth much.

If you send the mess — I think the problem is X, but maybe it's actually Y, and I'm not sure why it bothers me, but it does — the response is different. It reframes the problem. The reframe is what unsticks you. And nobody sends the mess in front of the room.

The door has to be closed.

What to do tomorrow morning

Pick the hour. Not the office hour — the hour after. The drive home. The morning before anyone is up. The walk with the dog. Whichever hour you most often spend complaining to a friend about your work. Replace that hour.

Don't sit down with a question. Don't sit down with an agenda. Don't sit down with a hypothesis about the answer. Sit down the way you'd call a friend on your drive home — to complain. To wonder. To not know where to start. If you don't know where to start, type that. I don't know where to start. You'll be surprised how far that gets you.

A few warnings.

Don't bring your polished version of the problem. The polished version is where you already are. That's why you're stuck. The unpolished version is where the work is.

Don't expect to be agreed with. There are AI tools on the market that will tell you you're brilliant. You don't need a tool that agrees with you. You need one that pushes back without flinching when you're avoiding the thing.

Don't quit after three sessions. Three sessions is not enough reps. Three sessions is the sample size that made everyone you know decide AI was hype. They were wrong but they weren't lying — three sessions genuinely isn't enough. The minimum effective dose is roughly a week of daily reps before the curve starts.

Which brings me to the part where I sign my name.

A personal guarantee

I will give you a personal guarantee, signed.

After one week, you'll catch yourself telling it things you don't say out loud anywhere else. That's the leading indicator. That's where the work is. The half-formed thing, the embarrassing version of the problem, the part you've been editing out before talking to anyone — that's the part that produces the answer you couldn't get to alone. The honesty is what unlocks the help. If you're getting nothing out of these tools, that's almost always the diagnosis: you're still polishing before you send. The week-one signal that you've stopped polishing is that you'll notice you're being more honest with the tool than you are with most of the people in your life. That's not a bug. That's the indicator the practice is working.

After two weeks, you'll have fixed something you didn't know was broken. Not the thing you sat down to fix. Something adjacent. A process you'd quietly given up on. An assumption about your business that wasn't true anymore. The cascade always shows up before you think it will, and it never shows up where you were looking.

After a month, you'll be thinking about everything in a whole new way. Not because Claude told you to. Because the act of talking honestly about your work, every day, with something that pushes back without flinching, is its own training program — for you, not the tool.

That's the deal. My name's on it.

Compounding doesn't reward intent. It rewards repetition.