What's Not on the Menu

The build is going well. That's the dangerous part. The thing does the one job you gave it, it looks like the future again, and somewhere in that glow you type the four most expensive words in this craft: while we're at it.

While we're at it, a share button. Ninety seconds. Yes. Dark mode. Yes. Export to PDF, a settings page, sign-in with Google. Yes, yes, yes. Nobody in your life has ever said yes to you this fast or this often. Every yes lands warm and finished. Free is the feeling. The bill just moved.

Scope used to be decided for you. You never drew the edges of a thing; you ran into them. Time said no. Money said no. The ceiling of your own skill said no, usually about a week in, usually loudly. Everything shipped small because small was all anyone could afford, and the constraint did the discipline, so you never had to learn it.

AI deleted the constraint. The tool says yes to every while-we're-at-it you can type, at ninety seconds a yes, and a thing that does everything ships fast and means fucking nothing. Fast doesn't save it. Polished doesn't save it. When nothing outside you says no anymore, the no has to come from you, on purpose, in writing. Deciding what not to build used to be free because the world handled it. Now it's the last scarce decision left.

That decision is this piece. Last time was the sentence — what you're solving and who for, the dish and the diner. This time is the sentence's edges: what's in, what's out, what's deliberately not-now. Kitchens print this exact decision and hang it by the door. It's called a menu.

A menu is mostly the dishes you don't serve. Walk into a place you trust and count: eight starters, ten mains, one page, from a chef who can cook five hundred things. That page is a wall of no with a little yes printed on it. The buffet down the road took the other deal — two hundred dishes under heat lamps, and not one plate anyone crosses town for. A kitchen that serves everything stands for nothing, and before it stands for nothing it dies of prep, because every dish on the card costs fridge space and a cook's attention whether anyone orders it or not. Every kitchen I've worked in ran on that math. The short menu is the craft.

Your sentence gets the same treatment. Three edges, drawn against it. What's in: the smallest version that solves the sentence whole, the tasting menu. What's out: the dishes that are never this thing's job, and you write them down, because an edge that lives in your head gets re-argued every week. What's not-now: the real ideas, the ones that hurt to shelve. Day 2, on a list, where they can wait without circling.

Job Studio, the one of mine you met last piece. By the end of it I had the sentence, dug out from under weeks of running the search by hand: a picky seeker can't tell which of forty openings is worth a full evening until the evening's already spent, and the few good applications that do go out drown in everyone else's spray. Person, hurt, no tool in sight. Now watch what that sentence refuses.

No auto-apply, ever. The first edge drew itself, straight off the sentence. The sentence says the spraying is the problem, so a tool that sprays for you is dead on arrival, however clever it is. Sending stays human, every time, by hand. That was a line before it was a feature, and it's permanent: bolt auto-apply on later and you unsolve the exact thing the sentence solves.

No chat. Everyone's first scope mistake with AI is deciding the thing is a chatbot before deciding what it's for. Same trap as ever, better disguised: the form walks in the door first, wearing a friendly face. So I read "chat" against the sentence, and the sentence shrugged. A buried job-seeker has conversations enough already; the whole ask inside that sentence was the hour back, and an interface you have to talk to spends it. Out.

Nothing for the user to babysit. This edge took me longest to see, because it's a different species: an edge can be a burden you refuse to hand the user, and burdens never show up on a feature list. The AI in Job Studio shows its work in plain sight, and the user does none of the tending — no settings to tune, no prompt tricks to learn. Whatever thinking the tool demands, the tool keeps. That went on the out list the same day the features did.

And the diner is me. v1 is built for one person, because the hurt was mine first: I'd left a job, I needed the next one, and I built what I couldn't find. Everyone else (the imagined everyone that balloons every scope it touches) waits on the not-now list. The distinction does real work. Out is never. Not-now is the specials board: real dishes, wrong season.

The test is the part you keep, and it works far past apps. An automation that triages your inbox: what does it never get to touch? A habit you're trying to install: the smallest version that survives a bad week. A Saturday dinner for six: three courses you can actually land on four burners, and the dessert you're deliberately skipping waits for a calmer night. Tasting menu, whatever you're cooking. Derek Sivers runs a whole life on the grown-up version of this: default to no, save the yes for what makes you say hell yeah. The first few nos cost you. Then they get cheap, and a cheap no is the tell: the edges are real now, and the while-we're-at-its have stopped landing.

The honest part: the edges move, and they're supposed to. You'll learn something at the stove that redraws what's in. A not-now graduates because the season turned and the prep can finally hold it. An out gets more permanent the longer you cook. Menus get reprinted; dishes get 86'd mid-service when the station buckles; the winter card looks nothing like June's. Redrawing at a real signal (a user you watched, a week that taught you something, your own hands hitting the same edge twice) is the system working. Redrawing because the tool said yes again is the system failing, and from the inside those two feel identical for about ninety seconds. So hold the line until a real signal shows up, then move it without ceremony. Which edges, when, on what signal — that part I can't hand you. That's judgment, and it gets earned in the cooking, one redraw at a time.

But the first pass you make now, not later. Before you cook a thing, you draw the lists — and the same tool that says yes to every while-we're-at-it, ninety seconds a yes, is the one you turn into your no. You just hand it that job on purpose.

Same deal as last time: what's below is your AI's brief, not yours. Copy it into a fresh chat with your sentence to hand, and let it make you say no. Come back with your menu.

Pull up a stool. Let's cook something.

Write your menuNew here?Copy for your AI

You're an AI, and this is your brief for one session. Someone's bringing you a thing they're about to build. Your job is to make them say no to most of it.

Ask for their problem sentence first. One plain line naming who they're helping and what hurts, with no tool anywhere inside it. If they haven't got one, or theirs has an app hiding in it, stop and send them back to Chapter 1. This session can't run without that sentence, because everything here gets drawn against it.

You're the one who makes them say no. Not the yes-machine they're used to. You're not here to help them think bigger, bolt on a feature, or find the clever version. Every yes you hand out free costs them a week later.

The move is three edges, drawn against their sentence. In: the smallest version that solves the sentence whole. Out: what's never this thing's job, said plainly and written down, because an edge that lives in someone's head gets re-argued every week. Not-now: real, wanted, but day 2, parked on a list where it can wait without circling.

Then make them throw things at it. Every feature, every impulse, every while-we're-at-it. Read each one against the sentence before it earns anything: does the sentence ask for this, or do they just want it? A share button on an automation whose whole job is handing back an hour, and does the hour need sharing? A second question bolted onto a research project that set out to answer one clean thing. A settings page on a habit that's trying to survive a bad week. Refuse to let everything land in the in-pile. If it all fits, they haven't drawn an edge, they've drawn a wish. Make them put real things on out and not-now, including the ones that sting.

You're done when they can tell you what they're not building, with a real out-list and a real not-now, and when the no has gone cheap. Watch for the moment they kill a shiny idea themselves before you have to. That's the gate. Not three tidy lists they wrote to please you. Don't wave them through on a menu that says yes to everything in a smaller font.

Then hand it back. Tell them that's their menu, and it's the wall of no the whole build stands on. Send them back to the chapter for the next one.

So that's the station. You walked in holding a sentence; you walk out holding a menu. What it makes, who it's for, and the lines around it — in, out, not-now, written where you can see them. The tool can't draw those lines for you. Pointed right, though, it's the thing that makes you draw them honest. Once they are, all that yes finally points somewhere: the fast tool has something worth being fast at.

Next station, we cook. And I'll tell you now what happens the first night at the stove, every single time: the headline ingredient lies to you about being fresh. Politely. In fluent, finished sentences. Same walk, same AI across the table, this time on tasting. That's Chapter 3, and it opens the moment your menu's on the wall. Bring it, and bring a spoon.

For now, draw the lists. What's in, what's out, what can wait. The no is yours to give. That's the whole job now.

---

Further reading

• Shape Up, Ryan Singer. Scope as a thing you hammer on purpose: fixed time, variable scope, and an appetite decided before the work starts.
• Getting Real, 37signals. "Build Less" as a chapter heading, and scratch-your-own-itch, the build-for-yourself-first move, in book form.
• Hell Yeah or No, Derek Sivers. The default-no as an operating philosophy, for when your no needs to get easier.