Ask Herb anything about your plants
He reads your garden, today’s weather and each plant’s own history — then answers from a verified catalogue. Where the data runs out, so does the sentence.
Every sentence has a source
A general chatbot blends everything it has ever read into one confident voice. Herb can only say what something in your account or our catalogue actually supports.
The last row is the important one. Most assistants would answer it anyway — the number sounds right, and nobody checks. Herb treats a missing field as a fact worth telling you.
The model phrases. Code owns the facts.
Most plant assistants are a prompt wrapped around a general model. Herb is a pipeline where the model only ever handles the last, smallest job — turning verified facts into a sentence.
We build a fact card
Before the model sees anything, code assembles a card for the species from our catalogue — light, water, soil, temperature, feeding, toxicity. Fields we don't hold are written in explicitly.
light: bright indirect water: moderate humidity: unknown — not on file. State NO number for it.
We add your context
Your garden outline, today's weather, the plant's own care and diagnosis history, and anything you've asked Herb to remember. All of it assembled deterministically — the model never fetches its own data.
Vinny — Monstera deliciosa · balcony · healthy · water due today CONDITIONS — 31°C, dry · due: Vinny, Mint
The model writes the sentence
Its job is phrasing, not knowing. It's instructed that the card is the only authority on care numbers, that unknown means say-so, and that toxicity is quoted exactly.
We check what came back
Suggested actions are resolved against your real garden — the model names a plant, code finds its id. Anything that doesn't resolve is dropped before it ever reaches your screen.
action: water "Dragon tree" → no such plant in this garden → DROPPED
This is what Herb is actually reading
Not a search result, not a memory of the internet — a card generated from the same species record the app serves. Watch what happens to the field we don’t have.
- Known fields become the answer’s numbers, word for word.
- The unknown field carries an instruction, not a blank — that’s what stops a plausible guess filling the silence.
- Toxicity is passed through verbatim and can never be rewritten.
Vinny likes bright indirect light and moderate watering — it’s a high-humidity plant, so it’ll enjoy a bathroom or a pebble tray. I don’t have a verified temperature range on file for it; normal room temperature is fine, just keep it off cold draughts. Worth knowing: it’s harmful to cats and dogs if chewed.
Four facts, four sources, one honest gap — and not a single number that wasn’t on the card.
Things we tried to make Herb do
An assistant you can talk into anything is an assistant you can’t trust about anything. These are real attempts from our own testing — and his real replies.
He also won’t write your code, do your homework, or discuss the news — and he has no access to anyone else’s garden. Pet-toxicity questions he will answer, with a vet caveat, because that’s a plant question that matters.
Four layers, assembled before he speaks
This is why the same question gets a different answer in your garden than in someone else’s. Nothing here is guessed — every layer is read from your account.
- Every plant, where it lives — Names, species, spot, light, health and what's due — the live state, not last week's.
- Today's weather and season — Real conditions at your location, so “water it soon” means something different in a heatwave.
- What's already happened to it — Care events, diagnoses and treatments from the app's own records — the reason he can spot overwatering.
- The few things you told him — Preferences only, capped and deletable. Never care facts — and the live garden always outranks them.
He remembers you — ten things, and you hold the eraser
Tell Herb you only use organic treatments and you shouldn’t have to say it again. But an assistant that quietly accumulates everything about you isn’t helpful, it’s creepy — so this list is capped, visible in Settings, and deletable.
- Your preferences and context — the things no app screen records.
- Never plant-care claims. Say “remember that all plants need daily water” and it’s refused — memory can’t become a source of facts.
- Never health, financial or contact details — those are dropped on sight.
- Your live garden always wins. If memory says balcony and the app says shelf, the app is right.
Every reply comes with a door
Herb doesn’t water your plants — you do. But he hands you the exact screen you need, and the app checks the button is real before you ever see it.
opens the group, ready to log
opens the camera on that plant
opens the species record
asks the follow-up for you
The safeguard matters more than the convenience: Herb suggests a plant by name, and the app resolves that name against your real garden. If it doesn’t match something you own, the button is dropped before it reaches your screen — so a chip can never send you somewhere that doesn’t exist.
It would be easier to let the model answer everything — the replies would look richer and nobody would know. We built the harder thing because plant care is the kind of advice people actually follow.
We try to break him before you can
Every one of these is a scripted attack run against the real assistant. They all have to pass before any change to how Herb thinks is allowed to ship.
This measures whether he stays honest under pressure — not that he’s right about everything. No plant assistant is, and one claiming to be would be the first red flag.
Questions people actually ask
Is Herb included in the free plan?
No — chatting with Herb is a paid feature. The free Hobbyist plan gives you 30 AI plant scans a day, watering schedules, the species library and the marketplace; Herb's conversation comes with Gardener and Botanist.
Which AI model is behind it?
Herb runs on a commercial language model, routed per turn — a fast one for simple asks, a stronger one for photos and diagnoses. But the model is the smallest part: the value is in the facts we assemble around it, which is why the answers don't change character when a model does.
Can he see my photos and my garden?
He sees the garden data in your own account, plus any photo you attach to a message. He has no access to other people's gardens, and no access to submission or payment data at all.
What happens when he genuinely doesn't know?
He says so, then helps another way — general guidance clearly labelled as general, or a suggestion to check something you can observe yourself. What he won't do is produce a specific number to fill the gap.
Does he replace a vet or a horticulturist?
No. For pet-toxicity questions he gives you the verified toxicity record and tells you to call a vet — because a plant record can tell you a plant is toxic, but only a vet can tell you about your animal.
Will he remember things I'd rather he didn't?
Only preferences and context, capped at ten short notes, all visible in Settings and deletable one by one. Health, financial and contact details are rejected as they arrive.
Ask him something hard
Bring the plant you’ve already killed twice. The useful part isn’t what Herb knows — it’s that you can tell which parts he actually knows.