
(Overview)Deme is a companion that knows your garden. This is the marketing site that has to find her an audience while my co‑founder builds the app: it reaches gardeners rather than software buyers, it lets you talk to her rather than read about her, and it had to do both without looking machine-made.
- Client
- GardenHarvestAI
- Sector
- AI products
- Year
- 2026
- Role
- Design and build
- Scale
- 11 pages · 18 guides
- Status
- Pre-launch
The colours, named for where they came from
- InkSoil#1A120A
- DemeSage#607A52
- SecondaryEarth#8B6B3D
- AccentTomato#C4521E
- GroundCream#F5F0E8
Type
Fraunces carries the headlines. Instrument Sans does the reading.
Gardeners are not shopping for software, and Deme is software.
The person this has to reach is standing in a garden wondering why their tomatoes split. They have never heard of Deme, they are not looking for an app, and the answer they want has nothing to do with software. A companion is also harder to sell than a tool. You cannot judge one from a list of features, because it only proves itself after a few weeks in your garden. So the site has to ask a stranger for trust before it can show them anything worth trusting. That is why it cannot open by explaining itself. It has to be useful first, to somebody who did not come looking for it.
It gets read in two places that could not be less alike: on a phone held in one hand outdoors, and on a wide screen at a desk. Most sites pick three or four sizes and let everything in between be somebody else’s problem. This one runs on a single continuous scale from a 320 pixel phone to a 2,600 pixel display, so every width in that range is a size it was drawn for rather than a size it survives. That is the sort of thing nobody thanks you for and everybody feels, and it only gets decided when the person drawing it is the person building it. That is most of what I am for.

An AI product cannot afford to look like one made it.
It started out illustrated with generated stock, and it showed. Waxy skin and airbrushed faces, the golden-hour glow on every outdoor shot, and files at 1344 by 768 because that is the size the generator makes rather than the size a 2,600 pixel layout needs.
The worse ones were wrong about their subject. An article on checking your broccoli for firmness ran a photograph of a cauliflower. Companion planting, which is the practice of growing things next to each other, ran a flat-lay of picked produce with oranges in it. The plant-diagnosis feature, whose whole job is recognising a sick plant, illustrated itself with a healthy one.
Thirty-five images were re-directed and replaced. Each one started by reading the thing it had to illustrate, the article or the section and its alt text, so the picture was briefed against the actual meaning rather than a vague mood. Then natural daylight with the glow explicitly forbidden, hands that are the right shape, and Australian specificity down to the Colorbond fences and the sugar-cane mulch, because the readers are in the southern hemisphere and the generator assumes otherwise.
The replacements are generated too, directed rather than photographed, and calling them anything else would be the same dishonesty moved to a different place. What is on show is art direction: killing the tells, insisting on local truth, and refusing a picture that contradicts the words beside it. Polish is table stakes. Making the picture mean the right thing is the work.

You cannot describe a companion. You can let somebody talk to her.
A feature list cannot carry a companion. Reading that Deme knows your garden is not the same as asking her why the lower leaves are yellowing and getting back something that sounds like a person who has grown tomatoes before.
So she is on the page rather than described on it. The hero opens with her instead of a screenshot of her, and the first thing you can do is type a question into it. She turns up again through the site wherever a claim is being made, so the evidence sits beside the assertion instead of a page away from it.
The frequently-asked page is where it earns its keep. A list of questions somebody else chose always misses yours, and the usual answer to that is a contact form and a wait. Here the list ends and Deme begins: if your question is not on it you ask her, and you get an answer in the place where you got stuck.
Her answers are curated for now. The model is not wired yet, and the build treats that as a launch item rather than a finished feature. What is built is everything around it: the placement, the fallback, a rate limit of five messages a day per visitor, and a seam the model drops into without changing one thing a visitor sees. The site is asking people to join a waitlist rather than to buy, and the honest way to do that is to let them try the thing they are waiting for.

One pair of hands only stays consistent if the consistency lives somewhere.
Forty-four components, none of them from a component library, on three layers of tokens: a raw value becomes a named colour becomes a component setting, so nothing reaches past the layer above it and the whole palette moves from one place.
One class on a section re-declares every semantic colour to the dark set, so a whole subtree changes theme without touching a single child inside it.
It is written down across 2,211 lines, every token group and all forty-four components with their usage, and there is a page that renders the lot at production width with the interactive ones actually working. Not a file of intentions. The built system, running.
This is the part that makes the rest possible. Eleven pages and eighteen guides stay consistent because the system makes the inconsistent version the harder one to write, and a scale that holds from 320 pixels to 2,600 is only maintainable as a formula rather than a decision taken again on every page. The documentation is the part that survives me leaving.

The useful part is at the front and the ask is at the foot.
Eleven pages and forty-four components, designed, built, written and art-directed by one person. What I would want judged is not the volume, it is the judgement, and the same trade turns up three times.
Leading with the product would have been quicker. Writing eighteen useful things first is the only way that survives a reader working out what you are doing. Replacing thirty-five images nobody had complained about is the same trade. So is building Deme into the page instead of describing her.
Each time, the expensive option had a visible cost and the cheap one had an invisible one. That route needs one person who can set the strategy, draw it, build it, direct the pictures and then write the eighteen useful things. It is the writing that kills it: the plan survives the meeting and dies at the blank page. Nothing here is measured, because nothing has launched yet.
- Days, start to finish
- 9
- Smallest to largest
- 8×
- Guides before the ask
- 18
- Images re‑directed
- 35

I designed, built, wrote and art-directed the marketing site. Dave built the app it is selling.
What I did
- The token system behind it
- Forty-four components, no component library
- Every guide, written by me
- Thirty-five images, art-directed
- Deme, placed across the site
What the team did
- Dave built the app
- The model, once it is connected
- The web app UX, which I am on now
01 / 06
Use the left and right arrow keys to move through the images.
(The call)Eighteen guides about growing things go up before a single line about the product, because a gardener will read one and would skip the other.
Leading with the product is what almost every site of this kind does, and its cost is invisible: the reader who leaves in two seconds never turns up in any number, so the version that loses them looks perfectly healthy. This way round the cost is visible and it is large. Eighteen guides is weeks of writing before a single visitor arrives, and if they are not genuinely useful there is nothing else on the page to catch anybody.
The site








Teaching first and not looking machine‑made were one job, not two.
Give the reader something before you ask them for anything. It is the oldest rule there is and the one every product site breaks first.





