📝 1. Use Obsidian to Improve Your Markdown Editing Experience
Editing .md files directly inside Claude Cowork is painful, inefficient, and prone to burning large amounts of tokens on mistakes. The author recommends downloading the free Obsidian app and linking it to your Cowork workspace folder. With Obsidian you can visually preview and edit Markdown formatting, which improves readability. Locking reading mode and enabling display of non-Markdown files like PDFs and images makes managing config files and memory banks feel professional and effortless.
Bloated root files generate massive redundant token consumption on every conversation, degrading model performance. The author proposes the "300-line rule": keep only core instructions in the root file (memory system, preferences, rules, routing map) and move non-essential rules out, referencing them with pointers. This on-demand loading strategy for task-specific rules shortens the total file length and measurably cuts token costs by around 25%, keeping the AI running in its leanest and most efficient state.
🧠 3. Scientifically Separate and Organize Memories vs. Rules
Many users conflate Claude config files with memory banks, causing output quality to drop. The author recommends a two-part test to tell them apart: anything that is an instructional behavior rule (using words like "always" or "never") belongs in the Claude config file, while factual-state information that may change over time belongs in the memory bank. Periodically having Claude self-audit and categorize these entries keeps the workspace clean. Pairing this with a "memory diet" strategy — setting a 150-line cap and building an automated archiving system — preserves long-term value without adding to the token load.
🚀 4. Migrate from Claude Projects to the Cowork System
Claude Projects are convenient but come with interaction limitations: you can't directly edit instructions or have the AI read and write documents on its own. The author demonstrates a one-step migration into the Cowork architecture: project instructions become workstation config files, project memories become memory bank files, and resource folders manage knowledge documents. This migration lets the AI manage documents directly, and by automatically adding routing entries to the root map, enables seamless switching between workstations so your workspace keeps evolving with use.
🛠️ 5. Know the Difference Between a Workstation and a Skill
Many users aren't sure when to create a Workstation versus when to write a Skill. The author identifies the core distinction: "Is this an ongoing domain?" versus "Is this a repeatable action?" A Workstation is a persistent work environment with its own memory, knowledge context, and tone — suited for long-term projects. A Skill is a set of instructions built for specific, standardized, repeatable tasks. By handing tasks requiring human judgment to a Workstation and fixed-process tasks to a Skill, users achieve deeper automation and upgrade AI from a helper tool to a high-efficiency autonomous executor.
Highlights
📝 Use Obsidian (free) linked to your Cowork workspace folder to visually preview and edit Markdown files — this alone eliminates a major source of token waste from formatting errors.
📏 Apply the 300-line rule: keep only core routing logic in your root file and load domain-specific rules on demand, cutting token costs by roughly 25% per conversation.
🧠 Separate rules from memories using a two-part test: instructional behaviors ("always"/"never") go in Claude config; time-sensitive facts go in the memory bank — mixing them degrades output quality.
🔄 Migrate from Claude Projects to Cowork by converting project instructions into workstation configs and project memories into memory bank files, enabling the AI to read and write documents autonomously.
🛠️ Choose Workstations for ongoing domains with evolving context and memory; choose Skills for fixed, repeatable, standardized processes — using the wrong type is the most common setup mistake.
⏳ Set a 150-line memory cap with automated archiving to preserve long-term value without bloating token load — a system that compounds without growing expensive.
