About
Seedwork
Seedwork is a methodology for building software with Large Language Models (LLMs)—systems with emergent capabilities that do things you didn't explicitly program.
The name comes from the observation that working with LLMs is more like gardening than construction. Capabilities aren't built; conditions for them to emerge are cultivated. The potential exists in the model before a line of code is written.
The term "Seedwork" has roots in software development, introduced by Michael Feathers and Martin Fowler. See Fowler's 2003 article. They used it to describe minimal, adaptable code—as opposed to rigid frameworks. As Fowler put it: "A framework is supposed to be a part-baked application that you extend in controlled ways to provide what you need. A seedwork is some minimal functionality that you modify however you like to get what you need."
The concept took hold in the domain-driven design community, where seedwork patterns provide reusable building blocks that developers adapt to their needs. See implementations in .NET and Python.
We use the term here for a related idea: a methodology for cultivating AI capabilities rather than constructing them. The intuition is the same—start minimal, adapt, let it grow organically. The application is different: not code patterns, but a way of thinking about building with systems that surprise you.
The Seedwork Manifesto was written by Michael Knapp in collaboration with Claude (Anthropic) in December 2024. The ideas behind Seedwork were shaped by many conversations with software engineer Ben Tucker.
Michael Knapp
Michael Knapp builds technology for mission-driven organizations. He has spent the past year exploring how AI can create value for nonprofits and social enterprises—and learning that the most valuable outcomes are often the ones you didn't plan for.
Michael is founder of sdoh.org and beyondhate.com. He recently joined the board of Native Plant Trust, which seemed relevant.
Contact
For questions about Seedwork or collaboration inquiries, reach out via LinkedIn.