The suggestion that names your next cabinet does not come from a language model. It comes from the same engine that draws a whole kitchen — run one step at a time, over the room you have actually drawn.
Ten cabinets in, the room stops telling you what it wants. The sink base is under the window because that is where the plumbing already is, the fridge is in the only corner that will take it, and wall two is finished and looks right. Wall three is a stretch of empty floor and a decision you have been circling for twenty minutes — and the honest reason you are circling it is that you cannot tell whether the next cabinet is a choice or a consequence. That is the exact gap the step planner exists to fill, and this chapter is about what it does in there.
A blank room has an obvious first move: draw the walls. A finished kitchen needs no move at all — you are reading the quote and arguing about handles. Almost nobody gets stuck at either end. They get stuck in the middle, with four or ten items placed, in the state where the plan is real enough to be worth protecting and unfinished enough that the next move might undo it.
People arrive at that point carrying three questions, and usually ask only the first one out loud.
The step planner answers the first question properly and helps with the third. It deliberately does not answer the second — it will tell you what fits next, never whether the kitchen you are building is a good one. That distinction runs through the rest of this chapter, and it is the honest boundary of the feature rather than a gap we intend to close.
What they search at midnight “what cabinet goes where” — almost always typed with a half-drawn kitchen open in the next tab, at the point where the next move stopped being obvious.
This is the single fact worth taking away from the page. When you ask what to do next, the question does not go to a language model. It goes to the planner’s own step logic, running on the server, over your actual layout — the walls you drew and their lengths, every cabinet and appliance you have placed, and where each one sits. It is the same engine that produces the full automatic design, stopped after one move.
Why that matters is easiest to see by imagining the other build. A language model asked for a next step will produce one. It will produce one at two in the morning for a wall that is 41″ long with 38″ of cabinet already on it, and it will sound exactly as assured as it does when it happens to be right. Fluency is not the signal you need. The only thing you actually want from a next-step answer is that whatever produced it knows the width of the gap.
An engine that measures the gap has two honest outcomes and no third. Either something in the catalogue fits there, in which case it can name it, or nothing does, in which case it has to say so or ask you something. Being told that a question has to be answered first is a better answer than a plausible cabinet that does not fit — because the plausible cabinet costs you the twenty minutes it takes to place it, look at it, and work out why the run no longer reaches the corner.
Ask the same question twice on the same layout and you get the same answer twice. Nothing is sampled, nothing is phrased differently on the second run, and there is no version of this where the recommendation changes because you asked in a different tone of voice. If the answer changes, your kitchen changed. That is a small promise, and it is the one that makes the suggestion usable as a working tool rather than a novelty.
The language model is still in the room — it is what reads the answer to you. Emma calls the step planner as one of her tools, gets back a structured result, and says it in plain language: “a thirty-inch base cabinet, B30, on wall three, left of the fridge.” What she does not do is invent the recommendation, improve it, or soften it. She is the voice on the answer, not the author of it. What she can and cannot see is a chapter of its own.
There is no fourth shape and no combination of two. Knowing the three in advance means you can tell instantly whether you have been given something to do, something to decide, or something to fix.
| Shape | What comes back | When you see it | What to do with it |
|---|---|---|---|
| Recommendation | One specific cabinet, which wall it goes on and where along it, plus the reason | The normal case, once there are walls and the next move is unambiguous | Agree and let her place it, place it yourself, or ignore it and place something else |
| Question | The one thing it needs settled before it can recommend anything — with two named options where there is a genuine choice | An empty room with no walls; a corner where a square and a blind cabinet are both reasonable | Answer it, then ask again. Working around a question just postpones it |
| Blocked | A statement that it cannot proceed, and the reason why | What you have already placed leaves nothing that fits | Undo, remove something, or accept the wall is finished and move to another one |
The empty-room case is worth saying out loud because it surprises people who expect a design to appear. With no walls yet, the answer is the question add walls first — not out of pedantry, but because there is nowhere to put a cabinet. A room that does not exist has no wall lengths, no corners and no constraints, and every recommendation the engine makes is a consequence of those three things. Draw the box first; it takes a couple of minutes.
One answer comes back at a time, and that is a decision rather than a limitation. A list of the next six cabinets is not a suggestion, it is a design — and a design generated around a half-finished room quietly commits you to five decisions you have not made yet. There is also a mechanical reason: each recommendation is computed against the room exactly as it stands, and the moment step one lands, that room is gone and steps two through six were calculated for a kitchen that no longer exists.
If you do want the whole thing at once, that is the full automatic design, which is a different feature and needs a free Google sign-in. The step planner is the one you use when you want to keep hold of the wheel.
On an L-shaped room the corner cabinet is the only box whose width comes out of two walls at once. Everything on both runs is positioned relative to where that corner ends, which makes it the constraint that sets the other two decisions rather than one item among many. So the planner works corner-first: on an L it recommends the corner cabinet before it recommends anything in the runs on either side.
Leave the corner until last and the failure mode is predictable. You fill both walls neatly from their far ends, they arrive at the corner from two directions, and what is left between them is a gap that no corner cabinet in the catalogue matches — too small for the one you wanted, too large to close with a filler you are happy to look at. Nothing in that plan is wrong individually. The order was wrong.
There is a second reason, which is that the corner is where the honest ambiguity lives.
A preference, not a fit problem — so it hands you both options by name.
When both are genuinely possible in your corner, the engine stops and asks, with the two options named, rather than choosing on your behalf and moving on. A square corner cabinet turns the corner properly and gives you a door at 45° or a pair meeting in the middle. A blind corner lets one run continue past the other, with part of the box tucked behind its neighbour — more usable width in the run, less convenient storage in the dark half unless you buy the pull-out gear.
That is a decision about how you want to reach into a corner for the next fifteen years, and nothing in the layout data settles it. A tool that quietly picked one would be making a lifestyle decision under cover of a geometry decision. The corner chapter goes through both properly.
Once a corner is settled it stays settled: the planner never recommends the same corner twice. Without that rule, a corner you deliberately left empty would come back on every single ask, and you would spend the session declining the same suggestion instead of finishing the wall.
When the engine runs, every cabinet and appliance already in your room is treated as fixed. It is not a starting suggestion it may revise, and it is not a draft it will tidy. It designs around your decisions instead of overruling them, which is a stronger promise than it sounds and the reason the feature is safe to use on a plan you care about.
What that buys you is the ability to start from the one thing you are certain about, which is almost always how real kitchens get planned.
The cost of locking is real and you should know it before it happens. If the cabinets you have placed make the rest impossible — a run committed in a way that leaves 9″ against a corner, say — the answer that comes back is that it is blocked, with the reason. It does not resolve the conflict by moving your work. We would rather hand you a blunt refusal than have you look up and find the sink base has migrated 6″ along the wall to make somebody’s algorithm comfortable. Watching a tool silently rearrange work you did with your own hands is the moment people stop trusting it, and they are right to.
Undo in the planner walks back the last three changes — deliberately, so that a large scene does not sit there eating the browser’s memory. Anything older than that cannot be undone. If you are about to pull four cabinets out to see whether the engine likes the wall better empty, save the design under a name first and experiment freely. Saving needs the same free sign-in as everything else that leaves your browser.
Here is the boundary, stated before you go looking for it: the step planner walks you through the floor cabinets — the base run and the tall run. Wall cabinets are not its job. They come from the full automatic design, or from you, with the mouse, once the floor is settled.
That sounds like an omission and it is closer to an order of operations. The base and tall run decide almost everything above them. They fix the countertop’s shape and its runs, where the sink lands, which openings the appliances get, and how the corner resolves — and the wall cabinets are then positioned relative to what is underneath. The cabinet above the cooking surface is not really a free choice at all: it is decided by the hood, and the hood is decided by where the range ended up in the base run. The clearance numbers behind that are in the appliances chapter.
Solve the run that constrains before the run that is constrained, and the wall cabinets become a short, easy job. Do it the other way round and you will be moving uppers twice.
We would rather write this down than let you discover it by asking six times on a wall with no floor cabinets left to recommend. If your base run is complete and you are waiting for it to start suggesting uppers, it is not going to — that is the automatic design, or your afternoon.
The step planner is one thing with two front doors, and the answer is identical through either. What differs is what it costs you to ask and what happens next.
By voice, asking what to do next is one of Emma’s twelve tools. Left-click the avatar in the corner of the canvas to start the call, say the sentence, and she reads the result back in plain language — catalogue codes said like a person, “a thirty-inch base cabinet, B30”, rather than spelled out like a barcode. Then she asks whether you want it placed. She places it only after you agree out loud, and when she does it lands as one step of undo, exactly like a cabinet you dropped with the mouse. Nothing else in the design moves.
The price of that door is the sign-in, the microphone permission, a browser that supports the call, and available time: ten free minutes a day per account, resetting on the Toronto calendar day, and one GH Coin per five-minute block after that. A block only renews if you actually spoke during it, and spoke within the last minute and a half — so leaving the call open while you think does not quietly consume anything.
On which note: if you go quiet for five minutes the connection closes on purpose. The microphone stays on, but nothing leaves your computer — the page is only measuring the level locally to notice you speaking again — and she comes back with the thread intact when you have been talking for a moment. Around twenty minutes of silence and she releases the microphone entirely.
Without a microphone, the same suggestion surfaces in the planner itself, so the answer is not gated behind a call. What you do not get on that route is the placement offer — you read the recommendation and place the cabinet yourself. One click, and the same single step of undo.
The full command reference is next door, and the walkthrough takes an empty room to a saved plan in one call.
None of these are difficult. All of them come from watching people get less out of the feature than it can give.
Which brings the chapter back to where it started. A suggestion tells you what fits next. It does not tell you whether the fridge door will clear the counter, whether two people can pass with the dishwasher open, or whether you have left yourself anywhere to put a hot pan down. Those are the questions that get expensive on delivery day, and they are a different job.
Before you order anything, read the twelve layout mistakes with your plan open beside you. The step planner and that checklist answer different questions, and you want both answered.
It runs the planner's own step logic over the layout you have drawn and returns one of exactly three things: a recommendation naming a specific cabinet, which wall it goes on and where, with the reason; a question it needs answered first; or a statement that it is blocked and why. One answer at a time, computed against the room as it currently stands.
No. The suggestion is computed by deterministic step logic running on the server over your actual layout, using the same engine as the full automatic design. Ask twice on the same layout and you get the same answer twice. The voice assistant reads the result aloud in plain language, but she does not invent it or change it.
Because something has to be settled before a recommendation is possible. With no walls drawn there is nowhere to put a cabinet, so the answer is to add walls first. At a corner where both a square corner cabinet and a blind corner cabinet are reasonable, the choice is a preference rather than a fit problem, so it hands you both options by name instead of picking for you.
On an L-shaped room the corner is the only cabinet whose width comes out of two walls at once, so it sets the position of everything in both runs. Fill both walls first and they meet at a corner that no catalogue corner cabinet matches. The planner settles the corner before the run, and it never recommends the same corner twice.
No. Everything you have already placed is treated as locked, and the engine designs around it. That lets you start with the things you are certain about, such as the sink base under the window or the fridge in its only possible corner. If your locked cabinets make the rest impossible, it tells you it is blocked and why, rather than quietly moving your work.
No, and this is the honest limit of the feature. The step planner walks you through the floor cabinets, meaning the base run and the tall run. Wall cabinets come from the full automatic design or from you. There is a reason for the order: the floor decides the countertop, the sink position, the appliance openings and the corner, and wall cabinets are positioned relative to what sits underneath them.
It means nothing in the catalogue fits the space that is left, given what you have already placed and locked. The message comes with the reason. The fix is on your side: undo a recent change, remove a cabinet that over-committed a run, or accept that the wall is finished and start asking about a different one.
Yes, on a voice call, and only after you agree out loud. A recommendation read aloud has changed nothing until you say yes. When she does place it, it lands as one step of undo exactly like a cabinet you dropped with the mouse, and nothing else in the design moves. The typing panel has no tools and cannot place anything.
No. It tells you what fits next, not whether the kitchen you are building is a good one. It has nothing to say about whether the fridge door clears the counter, whether two people can pass with the dishwasher open, or whether there is anywhere to put a hot pan down. Check those yourself against a layout checklist before you order anything.
Walls first, then the one cabinet you are certain about, then the sentence that does the most work in the whole planner.
Open the 3D planner →