Experimenting with intent cells for rerunnable LLM workflows

The author proposes "Intent-Cell Coding" (ICC DSL), a local-first notebook syntax designed to make multi-step LLM workflows reproducible by combining task text with execution details.

• ICC tracks which model or provider ran each step, the files or upstream outputs used, active constraints, produced artifacts, verifier results, and step dependencies.
• The approach aims to balance the exploratory nature of chat with the reproducibility of scripts.
• A key focus is recording exact input versions, hashes, snapshots, and artifact references to ensure reruns are comparable.

The author seeks community feedback on how others manage state for LLM demos and workflows, asking whether they use notebooks, app logs, traces, or config files.