Running Claude Code as an operations system: field notes

Most people meet Claude Code as a coding assistant: open a repo, ask it to fix a bug, review the diff. That's the front door, but spend enough time in it and a different shape shows up. The same primitives that let it edit code (a filesystem it can read and write, hooks that fire on its actions, a memory it carries between sessions) are enough to run the back office of a one-person operation. For the last few months that is what I have been doing, and I wrote up the patterns in an open handbook.

The core idea is that the task list lives in the repo as plain markdown files. Every open piece of work is a file with YAML frontmatter (status, priority, due date, which project it belongs to), read and written by a handful of small scripts. Because the tasks are plain files in git, the history of what I worked on is just the commit log, and Claude can plan my day by reading the same files I do. There is no API to call and no separate database to keep in sync.

Hooks make it stop feeling like a chat box. A PreToolUse hook can block an action before it runs; a PostToolUse hook can record what just happened. I use them to auto-commit finished work scoped to exactly the files a session touched, to guard against two terminals editing the same file at once, and to stop a class of mistakes I do not trust myself to never make. The agent proposes and the hooks enforce.

Memory closes the loop. A correction I give once gets promoted into a small markdown file that future sessions load on demand, so the same lesson does not have to be relearned. Over time the system accumulates the judgment calls instead of forgetting them at the end of each conversation.

It is a task store, some shell hooks, and a folder of notes. The interesting part is that wiring those three together turns a coding assistant into something closer to an operations layer that a small team can actually run on.

I collected the full set of patterns (task store, daily flow, hooks, session capture, memory, multi-terminal and multi-agent work) with real scrubbed examples in an MIT-licensed repo:

.