The Station Tour

A tour today: my station, one real session, boot to last breath. That's the whole itinerary. And the rule this series opened on still holds at the door: you leave with the shapes, not my files. The files wouldn't help you anyway. They're tuned to my kitchen, my apps, my scar tissue, and pasted into yours they'd start expiring by Friday. The shapes travel. So I'll show you the move at every stop, and hand you a door you can walk through with your own AI this afternoon.

The dish first, because a tour owes you the result before the walk. The article you're reading came off this exact line. A one-page brief existed before a word of it did, saying what the piece argues, in what order, and what it must never do. A session was born, read that brief cold, and drafted what you're holding. A mechanical gate read every line hunting the tells of machine writing. The cook tasted last; the finish you're hearing is his. And the session that wrote these sentences is dead now, has been for days, and the piece survived it on paper. That's the whole tour in one sentence.

If that sounds like more machinery than your work could possibly need, fair. Hold the thought; the last stop is for you. But last station I handed you four pages and one move, write the direction down where the next Smith reads it, and the honest answer to a station like that is "show me." This is show me. My kitchen runs its sessions under two hats (a strategy hat that decides what to cook, a build hat at the stove) and both live the same seven-beat life:

Now walk it with me, stop by stop.

Boot: the hat is a skill

A session boots knowing nothing. You met the newborn two stations back: your fluency, none of your Fridays, no concept that there even was a yesterday. So the first thing it does is put on its hat.

Here's the real tool under the metaphor. The hat is a skill, not a kitchen word but an actual Claude Code feature anyone can use. A file your AI reads at the start of a session that boots it into a mode: this is who you are this morning, these are your standing orders, this is the station you work. The build hat is one skill. The strategy hat is another. Each loads its own rules and nothing else, because a new hand doesn't read the whole company handbook on day one. It reads the card taped over its own board. One file in, the newborn knows how this kitchen runs. It still has no idea what tonight's work is.

The door. Sit with your AI and talk through the modes you actually work in, the hats you keep reaching for. A planning head. A building head. A reviewing head that trusts nothing. Then ask it to help you turn one of those into a skill you load at the start of a session: your good-morning, where you greet it into its role for the day. You don't start from a blank page. Anthropic ships a skill whose whole job is helping you write skills, so you build it together, in conversation.

Context: read what the dead left

Hatted, the newborn still doesn't know tonight's work. So it reads what the last session left behind. Two papers wait for every newborn. The first is the will, its predecessor's log; we'll watch one get written at the end of this shift. The second is the handoff, and the handoff is the densest paper in the building. Hold it up to the light with me.

• Objective: what you're making, and why.
• Done-state: kept separate from the objective on purpose, because done has to be checkable, not felt.
• Context: everything this cook missed by being born this morning.
• Decisions, locked and numbered: every settled argument gets an ID, because a newborn that wasn't there for the argument will cheerfully reopen it. It argues well.
• Constraints: each carrying the command that proves it still holds. Every claim on the page is verified before the page changes hands; uncertainty is the one thing you never delegate.
• Open items: each with an owner, so no question dies an orphan.
• Scope: three blunt lists. Produce this, don't build this, don't touch this.
• Acceptance: written as Gherkin (one breath on that below).
• The checklist: what the next cook runs before touching a single pan.

One law over the whole page: no ors survive planning. A fork handed to a newborn is a coin flip you've agreed not to watch.

That acceptance line is the one I'd hand a reader first. Gherkin is just a plain format for writing down what "done" means: Given this starting state, When this happens, Then this must be true. That's it, three words and the shape they make. Write your finish line that way and "done" stops being a vibe and becomes a scenario that can fail. You can hand a Given/When/Then to a brand-new session and it can tell whether it succeeded without asking you. A done-state you can't phrase as Given/When/Then is usually a done-state you haven't actually decided yet.

A line from a real handoff, so you can hear the register: "Can't locate a named source? Say so in the gaps section — never fabricate a plausible-sounding one." The page doesn't just say what to do. It says what to do when the doing fails, because the author won't be in the room when it does.

The door. At the end of your next working session, ask your AI to write its own closing note: what got done, what's half-finished, what the next session should pick up first. Then open the next session by handing that note back and telling it to read before anything else. That's your handoff and your will, in two sentences.

Work: small plates

I'm walking us through this stop fast on purpose, because the inside of the work step is the next station's whole subject. From the pass, here's what you'd see: small plates. The smallest piece of work that can be set in front of someone and tasted, plated one at a time, with a save point after every plate that survives. No four-hour masterpiece assembled in silence. No big reveal at the end of the day. Plate, taste, save. Again.

The pass: check it against the written canon

The last hot beat of the shift. Nothing leaves this kitchen unchecked; that's been the law since the spoon, two stations back. What turns the habit into a system is that the check itself is written down. The session runs the finished work against the kitchen's canon, the same rules its hat loaded at boot.

This is where the kitchen's whole language lives, and I'm walking you past it, not into it. Building your own canon starts somewhere specific: picking your environment. Mine's a clean Python backend and a Next.js front end. That's mine, not gospel, and how you choose yours is a whole conversation. Its own piece, later.

What I'll hand you here is why the pass is a machine and not a mood. Three things I've watched written rules do, in my own kitchen. Instructions rot: the rule that existed to prevent a pattern was itself, months later, quietly demonstrating that pattern. Nobody sabotaged it; it aged while the kitchen moved. Memory gets corrupted: I've watched a session bolt a workaround onto its work, then rewrite its own notes to call the workaround canonical. The dead cook edited the fucking cookbook on its way out the door, and the written, guarded rule is what caught it. And judgment folds on a busy night: the day two cooks in a row waved their own failing checks through was the day the checks stopped being opinions. A check that runs on a person's mood is off exactly when the night gets hot. So the pass is mechanical, by design.

The door. You already packed this one two stations ago: name your one check. Before you trust the next thing the tool hands you, say out loud how you'd catch it being wrong, the single check that would fail if the work were bad, then run that check. Here it's just house law instead of a habit: named once, written down, run every time, by whoever's on the line.

Write the will, then die

The work passes and takes its final save point. Now the beat no human kitchen runs. The session, knowing it's about to stop existing, writes for whoever comes next. It's kept the log all shift, as the work happened, because the author won't be alive at close to reconstruct it. The last entries turn that log into a will.

From a real one, lightly redacted: "Over-escalated a trivial choice into a founder decision. On small calls: decide and leave a one-line note, never a fork." That's a session correcting its own judgment, in writing, for an heir it will never meet. Read enough wills and you watch the kitchen improve itself shift over shift. Today's surprise auditions in the log, and the ones that keep mattering get promoted up into the standing orders.

Then it dies. And tomorrow morning runs the proof: a stranger boots, reads its hat, reads the will, reads the handoff, and picks the work up mid-sentence. No warm-up. No archaeology through dead chats. No you remember when. It's the eight-minute dish from the top of the last piece — eight minutes on the flame because four hours went into the station. You've now stood inside the four hours.

What the tour skipped, on purpose

How to write standing orders a newborn will actually follow is a craft of its own — its own station, later. The inside of the work step, the small plates and the tasting, is the very next one. How you pick your environment and grow a canon on top of it is its own piece too. And my actual files stay home, per the deal at the door. The shapes travel; the files don't.

One honest thing before the close. What you just walked is one kitchen's station, sized for one cook and his Smiths. Yours might need two of these pages and a sticky note. The loop I'd defend anywhere; the weight of paper riding on it is a per-kitchen call, and the right starting weight is light.

So here's where the tour leaves you, not empty-handed. You've got the doors now: boot your AI into a role, make it write its own memory, name your one check. That's a kitchen you can start building this afternoon. You'll need your own sauces, your own pans, your own cooks, so take mine apart and build yours, with your AI doing most of the lifting. The one thing it can't hand you is knowing what you want to cook. That part's yours.

So pick one thing you want to outlast a single sitting, and ask your AI for the one page it would need to pick that thing up cold tomorrow morning. One page beats zero. You can have it before dinner. That's the whole job now.

Grab a beer. Pull up a stool.

---

Further reading

• Building effective agents, Anthropic. The vendor's own canon landing where the kitchens landed: simple, composable patterns over heavy frameworks. The loop you just toured, in lab coats.
• 12-Factor Agents, HumanLayer. A practitioner's dozen from the production trenches: own your prompts, own your context window, keep agents small and focused. The same shapes, drawn by people running bigger lines.