I built a knowledge-work counterpart to my Claude Code starter prompt. The coding version manages a repo; this one manages a consulting project — an RfP response, a discovery, a research report — out of one shared folder that syncs to the team via OneDrive. The folder is the single source of truth and stays ahead of the chat.

Same hierarchical context management as the .md-file memory system, but adapted for work with no compiler to catch drift: a context/ layer of synthesized understanding, an inputs/ log where every email, doc and transcript is registered with an ID and source, and a working/ folder that moves through analysis → drafts → review → final. Three things code never forced: rigorous capture with source provenance, a hard split between what a source said and our interpretation of it (keep conflicts, mark "to triangulate"), and a client-safe vs. internal-notes distinction. Large inputs get digested by a subagent so the main thread stays clean.

Grab it: COWORK-PROJECT-OPERATING-MODEL.md. Fill the two brackets, say who can see the folder, and it scaffolds the rest.

Full write-up on why it's shaped this way: the context operating model for projects that don't fit in a chat.