CoolDeep AI Digest

The article introduces Claude Cowork, a desktop app that connects Anthropic’s Claude AI directly to the user’s actual files on their computer, allowing the AI to read, edit, and remember work across sessions. The author contrasts Cowork with the more familiar Claude Chat (a browser‑based question‑answer tool) and Claude Code (a developer‑focused terminal tool). While Chat requires users to re‑explain their project and manually copy‑paste context each time, Cowork loads the user’s workspace automatically, eliminating repetitive setup and saving significant time—estimated at 1.5 hours per week or roughly three full workdays per year. The core argument is that many users waste hours re‑introducing context to Claude each session, whereas Cowork provides a persistent “employee” that knows the user’s files, projects, and preferences. To help readers adopt Cowork effectively, the author shares five practical tips: 1. **Use Obsidian as a markdown viewer** – Cowork stores instructions and memory in .md files. Opening the workspace folder in Obsidian renders these files with clean headings, bold text, and formatting, making them easy to read and edit without learning Obsidian’s advanced features. Enabling “Show all file types” lets users see PDFs, images, and spreadsheets alongside markdown files in one sidebar. 2. **Keep CLAUDE.md under 300 lines** – This main instruction file loads every session; bloated files increase token usage and slow performance. Organize it into six sections: memory system, preferences, rules, routing map, references, and workstation creation. Use the routing map to load specific workstations only when needed, keeping the core file lean (200‑300 lines). One user reduced token consumption by ~25 % after trimming CLAUDE.md from 600 to under 250 lines. 3. **Put your memory on a diet** – Structure MEMORY.md into three sections: Active Projects (short status updates), Scheduled Tasks (recurring jobs), and Core Memory (stable personal facts). Enforce two rules in CLAUDE.md: each memory entry limited to one or two sentences, and a hard ceiling of 150 lines. When the limit is reached, Cowork compresses older entries into an ARCHIVE.md file that does not load at startup, preserving history without token cost. 4. **Migrate existing Claude Projects into Cowork** – Move project instructions to a workstation CLAUDE.md, project memory to a workstation MEMORY.md with clear sections, and knowledge files into a Resources folder. Create a blank .md file, paste the instructions and memory, save it, then ask Cowork to run the migration; it will automatically split content into the appropriate files, giving users full control over each component. 5. **Distinguish workstations from skills** – A workstation defines a repeatable process (e.g., “email workflow”) that loads specific instructions and memory when invoked. A skill is a discrete capability the AI can perform (e.g., “summarize a paragraph”). Understanding this prevents confusion and helps users organize their Cowork setup efficiently. Actionable takeaways: install Obsidian to view markdown files, trim CLAUDE.md to 200‑300 lines using the six‑section structure, enforce concise memory entries with a 150‑line cap and archive older notes, migrate any Claude Projects into Cowork workstations, and clearly separate workstations (process containers) from skills (individual abilities). Implementing these steps will reduce repetitive setup, lower token usage, and let Claude work as a persistent, context‑aware assistant.