Landkit: a folder template that runs a landscape project with an AI agent

Most landscape projects lose context somewhere between the third call with the client and the second revision of the concept. Photos pile up in a chat. The measured plan is on a laptop that someone took home for the weekend. The thing the client approved last month is buried in a thread that scrolled past two weeks ago.

We built Landkit to fix that for ourselves first. Then we put it on GitHub.

What it is

A folder. Just folders and markdown files. No software to install, no app to learn, no account.

The structure mirrors how a landscape project actually moves: admin and scope, site survey, measured base, client context, AI concepts, approval, documentation, estimate, final visuals, delivery. Each stage has its own folder, a short README, and a clear contract: what comes in, what goes out.

00_project_admin
01_site_survey
02_base_plan
03_client_context
04_ai_concepts
05_client_approval
06_design_documentation
07_estimate_and_phasing
08_final_visualization
09_delivery

You copy the template for each new client. That's the whole setup.

Why an AI agent

We run Landkit with Claude Code in the terminal (how that fits with the rest of our AI setup is in How we use AI beyond web chat). The first message to the agent is literally:

We are starting a new landscape project. Read START_HERE.md and guide me.

The agent reads the rules, looks at what's already there, asks one or two targeted questions, fills the right files, and tells you the next step. It doesn't try to design anything — that's still the designer's job. It just keeps the workflow clean so you can focus on the part that actually requires judgment.

We've been running real projects this way since early 2025. Plots from a few hundred square metres to three hectares. Russia, Europe, the UAE. Different service scopes. The folder structure held up across all of it.

What it is for

It's a working folder where photos, decisions, briefs, AI prompts, and final files all live in one place — readable by you, readable by your agent, readable by the next person who picks up the project six months from now. Drawing still happens in CAD, images still come from a generator, and the judgment stays with the designer; Landkit only keeps the record of all of it, and it works with any AI service on any operating system.

Open

Landkit is on GitHub, MIT licensed. Fork it, strip out what doesn't fit your studio, add your own pieces. The README explains how to adapt it for your own service scope.

If you try it on a real project, tell us what broke. That's how it gets better.