The Four-Layer Hierarchy for Claude Co-work Setup

The Four-Layer Architecture

To maintain AI performance as document volume grows, organize your workspace into four distinct layers. Skipping the first three layers leads to context bloat and degraded reasoning.

• Instructions (Layer 1): Maintain a claw.md file between 50 and 100 lines. Use it only for pointers to other files and critical "gotchas" learned from past errors. Avoid long, monolithic instruction files.
• Memory (Layer 2): Store persistent preferences (e.g., "emails under 200 words") and static facts (e.g., client names) in a dedicated memory file. Include instructions in your claw.md that allow the AI to update this file autonomously when you provide clear, new rules.
• Skills (Layer 3): Encapsulate repeatable processes into folders containing a skill.md file. Bind these skills to specific project folders rather than keeping them globally available to prevent the AI from becoming distracted by irrelevant capabilities.
• Connectors (Layer 4): Integrate external systems (Gmail, Drive, Calendar) only after the first three layers are functional. Embed connector calls within specific skills rather than invoking them ad-hoc to ensure the AI follows a structured, predictable workflow.

Implementation Strategy

• Map Files: Create a table-of-contents file (map.md) that lists subfolders and files. This allows the AI to reference relevant content without reading every file in a directory, which preserves its context window.
• Skill Creation: Run a task manually with the AI until the output meets your standards. Once proven, use a "skill creator" agent to formalize the process into a reusable skill, explicitly instructing it to remain input-agnostic.
• Trust Building: When first connecting external systems, manually approve every action. Gradually increase autonomy only after the AI demonstrates consistent performance across a diverse set of inputs.