The Plan — An AI-Directed Tissue Culture Lab
We’re building a home tissue culture lab where Claude — Anthropic’s AI — directs the experiments. I run the lab solo from my room in London. My project partner Zi contributes remotely from San Francisco. Claude acts as lab director: it designs protocols, instructs me step by step, I execute the physical work, and Claude evaluates outcomes from sensor data and my descriptions. Eventually Claude will assess photos directly — but initially, the feedback loop runs on numbers and text. AI proposes, human executes, AI learns from results.
This is not science fiction. LLMs have already been shown to autonomously design and optimise wet lab experiments. Carnegie Mellon’s Coscientist (Nature, 2023) had an LLM autonomously plan and execute chemistry experiments. OpenAI’s work with Red Queen Bio showed GPT-5 optimising a molecular cloning protocol across multiple rounds, achieving a 79x efficiency gain. We’re applying the same principle to plant tissue culture, using a home lab setup costing ~£434 instead of million-pound robotics.
The core idea
Claude acts as lab director. It designs protocols, provides step-by-step instructions, monitors the environment through sensors, diagnoses problems, and adjusts variables between iterations. Once a camera is connected, it’ll evaluate photos too. Every decision, observation, and adjustment is logged. This is the design-observe-correct loop described in the research papers, running in a home lab.
Critical strategy: flywheel first, plants later
We do NOT start with plants. We start with proving the Claude feedback loop works on simple, low-stakes lab tasks. Once the system is smooth, we introduce living material. If the feedback loop breaks down on something simple like media preparation, we’ll waste plants. Get the mechanics right first on things where failure is cheap.
The phases
| Phase | Goal | Timeline |
|---|---|---|
| 1 — Prove the loop | Lab setup, sterilisation validation, media preparation, contamination dry run | Days 1-3 |
| 2 — First plant material | Explant preparation, culture initiation, observation protocol | Days 4-6 |
| 3 — Ongoing | Daily monitoring, subculture, protocol adjustments, scale up | Weeks 2+ |
What makes this novel
No one has published on “LLM-directed home plant tissue culture.” It doesn’t exist yet — we’re building it.
What exists and is proven: LLMs controlling chemistry labs (Coscientist), LLMs optimising biology protocols (Red Queen Bio), computer vision for plant disease detection, and a thriving home tissue culture hobby community.
What we’re combining into something new:
- Claude as lab director — protocol design, troubleshooting, adaptation
- Photo-based contamination monitoring via Claude’s vision
- Iterative protocol optimisation — Claude analyses results, adjusts variables
- Applied to commercial plant propagation — landscaping, rare plants
- No robotics — human hands, AI brain
- No million-pound equipment — home lab with a still air box
- Accessible to anyone with a Claude subscription and basic supplies
The reading list
We compiled 16 resources across four categories:
Textbooks: Bhojwani & Dantu’s Plant Tissue Culture: An Introductory Text (Springer, 2013) for theory, and Park’s Plant Tissue Culture: Techniques and Experiments (Elsevier, 2021) for hands-on protocols.
AI-directed research: The Coscientist paper (Boiko et al., Nature 2023), the Red Queen Bio results (OpenAI, 2025), Agent Laboratory (Schmidgall et al., 2025), and several survey papers on autonomous scientific agents.
Computer vision: Mohanty et al.’s deep learning plant disease detection (99.35% accuracy on 54,306 images) — TC contamination detection is a simpler version of this problem.
Aroid protocols: Zhang, Chen & Henny’s Syngonium somatic embryogenesis paper (2006) is the key one — 86% of petiole explants formed embryos, 50-150 plantlets per petiole. That’s our target protocol.
What’s arriving this week
- Raspberry Pi 4 (8GB) — the lab server
- ESP32 + DHT22 — temperature and humidity sensor
- Pressure cooker — for sterilising media
- Still air box — the low-budget laminar flow hood
- MS medium and agar — on order
The sensor infrastructure goes up first. Once Claude can see the environment, we start the biology.