llm-do

llm-do is my framework for building reliable LLM applications through progressive stabilizing.

The project is currently on hold. I now plan to focus first on Commonplace, which is a more basic exploration: before stabilizing full LLM applications, I want to understand how LLM-assisted knowledge work can accumulate into durable, inspectable structure.

The idea is to start with flexible LLM behavior, then progressively move stable parts into more constrained forms: prompts with clearer contracts, typed interfaces, tools, tests, and ordinary code. The boundary between neural and symbolic components should be movable instead of fixed up front.

Commonplace generalizes part of that lesson to knowledge work: an agent can start with loose notes, then progressively stabilize them into typed, validated, linked artifacts.