Farmscribe turns a farmer's field notes into structured records and advice that knows the farm. It started as a buried line in a Jira ticket. It's now a live app, in invite-only beta at my.farmscribe.app.
Visit the live app ↗
At LiteFarm, an open-source farm platform, I'd scoped a Farm Notes ticket down to a simple digital noticeboard. Buried in the "for later" section was one throwaway line: AI could auto-create tasks from notes. It never shipped there, LiteFarm had no AI budget for it. It became Farmscribe instead: a mobile note taker, built AI-first with a sustainable-agriculture domain partner, that turns a quick field observation into a structured record and advice grounded in the farm's own history. What follows is the short version of how it got from that line to a live app now in invite-only beta.
LiteFarm's constraints were real: no AI team, no AI budget, a careful governance process. The idea was right, the address was wrong. Spinning it out wasn't rebellion, it was the only way to build AI-first without asking permission for every decision.
"As a user I want to be able to share some observations or take quick notes around the farm"
As requested by multiple users, this is the feature for a separate section on the sidebar that works like a farm noticeboard. The minimizable sidebar could exist alongside the new planned insights dashboard. Please use existing free libraries wherever relevant to avoid building from scratch.
Eg: Broken fence, Goat Dolly due for vaccination next Thur, Coyote sightings in the area - keep the goats locked after 5pm, Researcher Dave visiting Mon, Fill fuel
These may or may not be actionable, so they need not necessarily have an owner or a due date.
Must have: It is a simple note publish feature with a simple long text, upload media and publish. It is largely like a whiteboard one might have in their barn or shed, with an additional ability to add notes to self. This is not intended as a 1-1 chat, nor a task, but we might potentially expand it based on how the feature is working out.
Enabled functions: to begin with: For author, delete. For others, just view.
Initially, the notes can just be private (note to self) or public (like a noticeboard in barn). All published notes must appear in reverse chronological order along with the author's name and date.
To begin with, we can support historical notes up to a month, with archiving possible for up to 6 months in the past. If the notes are beyond 25MB or over a month, we encourage owner to archive to be able to continue using the notice board.
Open point: Not sure if it makes sense to associate a field / location to a note?
Great to have: When someone publishes a note to the noticeboard, a notification goes to all the members of the farm. Ability to delete posted note. Members can react with basic few emojis. A note can also be accompanied by one picture (jpeg / png). On mobile, can upload or open camera.
Unsure of: download content of noticeboard.
Success criteria: %Usage = Number of farms using the farm noticeboard / Total number of true farms. If the % of farms using noticeboard is ≥15% of active farms, then the feature is considered successful.
For later scope: Based on user feedback, we can plan for specifically tagging people, creating tasks or custom tasks from notes. Depending on success of the feature, we could have AI join the chat and auto-create custom tasks from notes. For example, someone notices broken fence, AI can create a custom task "Fix fence". Possibility of moderation of posted content. Possibility to respond 1-1.
"Depending on success of the feature, we could have AI join the chat and auto-create custom tasks from notes. For example, someone notices a broken fence, AI creates a custom task 'Fix fence'."
The problem wasn't a missing noticeboard. It was that a farmer's best observations evaporate before they're useful. Aphids on the carrot bed, mentioned once, forgotten by the time they're an infestation. The fix isn't just logging, it's logging that gives something back. Tell the app what you saw, it structures the note and hands back advice. Logging stops being a chore and starts being a conversation.
Four stages. One principle: earn the right to add complexity before adding it.
Twenty-two words in a "for later" section: AI could auto-create tasks from notes. Not a spec, a throwaway thought. But it named the real gap, what a farmer writes down and what they actually mean to do about it. The question wasn't whether AI could do it, that's one API call. It was where the advice should live. During capture, or after? Blocking, or additive? The answer: after, never during. The farmer logs an observation in under ten seconds. The AI responds when it's ready. Action first, intelligence second. That sequence is the whole product.
Farmscribe is an AI product, built with an AI assistant. Worth being honest about. Claude was fast and genuinely useful for the structural work: specs, edge cases, drafts. Every real design call, where advice should live, how to build trust with a skeptical farmer, needed fifteen years of designing for field workers. No prompt carries that. The same gap exists in the product itself. A model can't know if its advice fits this farm, this season, this soil, unless it's fed that context. That's not a bigger model, it's design: the farm's profile, its history, the farmer's own corrections, the region's rules. Context engineering, not fine-tuning, entirely human craft.
The full loop runs end to end at my.farmscribe.app, voice, text, or photo capture; advice grounded in your own history; tasks the AI parses from a sentence; a farm map of drawn fields and dropped pins. Four languages, real accounts, a real backend. Invite-only, and real.
Log the same aphids twice and it connects them. Correct a priority and it learns. It won't recommend an input that's banned where you farm. That's what separates Farmscribe from a wrapper around a model.
The Open Furrow mark, a documented brand voice, and a landing page that shows the app running in four languages instead of pitching it, with pricing that adjusts to the visitor's region.
Voice, tasks, and a required farm location were all pulled in from "later" once the reasons became clear. The reversals stayed in the record instead of getting smoothed over.
The calls that mattered, and what changed my mind. The reversals are kept on purpose, in an R&D project the changes of mind are the work.
Instead of: Remain a feature inside LiteFarm
LiteFarm's resource constraints blocked the AI layer; going standalone removes the ceiling.
Instead of: AI during compose, blocking the save
Preserves barn-whiteboard speed; the AI is additive, never a gate.
Instead of: Text + photo only, voice deferred to Phase 2
Field-first means hands full of soil, voice is the fast lane, not a nicety.
Instead of: Task management out of scope for MVP
A note you commit to acting on is a task; collapsing the two kills orphan tasks.
Instead of: Location optional / skippable
It's the backbone for location-aware advice and the field map, at the cost of new tension with offline-first.
Instead of: Firebase, or a hand-rolled Postgres and auth layer on Netlify
The farm/membership/notes/tasks data is inherently relational; row- level security lets that access model live as declarative SQL policy instead of scattered application checks.
Instead of: Fine-tune a model per farm
A single farm's data is far too thin to fine-tune, and fine-tuning teaches tone, not a farm's facts; context engineering fits the existing seam.
Instead of: Keep the provisional skin, or reach for a generic agtech leaf
The mark should say what the product does, capture in and advice back, with no leaf and no soil-free AI-app look.
The most interesting product idea in a ticket is rarely the headline feature. It is the throwaway thought in "for later" that nobody had time to develop. Train yourself to find it.
AI that gates actions loses trust fast. AI that responds to completed actions builds it slowly. The sequence matters as much as the capability. Never block the user while the model thinks.
When an AI product feels like it knows your world, that is almost never a bigger or retrained model. It is the right context, your history, your corrections, your constraints, assembled and handed to the model at the moment it answers. Design the context, not the weights.
Fifteen years designing for people with no time for bad software: farmers, field workers, NGO staff in places where a confusing flow is a real problem. Farmscribe is what happens when that experience meets an AI layer built to respect the user's time instead of consuming it. It's live now, in invite-only beta, at my.farmscribe.app.