Most people use an AI assistant as a blank chat box. Every session starts from zero.
I wanted something more. I wanted an assistant that already knows how I work when I open it.
So I built a working environment around Claude. It lives in one folder on my Mac and has four parts.
- Operating manual (
CLAUDE.md). One file Claude reads at the start of every session. It sets my output format, my session modes, the thinking frameworks I want named, and my red lines. - Skills library. 200+ skills. Each skill is a written procedure for one kind of job, such as publishing this site or reviewing a plan. Claude loads a skill only when the task needs it.
- File-based memory. Small memory files with an index. Claude reads the index at session start and opens a file only when it is relevant. A Notion database mirrors the memory so I can browse its history.
- Subagent routing (
SUBAGENTS.md). Rules for when to hand bulk or parallel work to a helper agent. This keeps the main conversation focused.
Each part fixes one failure of a plain chat. The manual fixes inconsistency. Skills fix repeated instructions. Memory fixes amnesia. Routing fixes clutter.
The whole folder is versioned and backed up every night by my memory and skills backup system.