Unknown Title

Key Concepts

• Context Window: The limited amount of information (tokens) an AI model can process at one time.
• Claude Code: A command-line interface tool for interacting with Claude.
• claw.md: A configuration file used to provide instructions and project context to Claude.
• MCP (Model Context Protocol): A standard for connecting AI models to external data sources and tools.
• Sub-agents: Independent AI instances spawned to handle specific, siloed tasks to prevent context bloat.
• Context Compaction: The process of summarizing conversation history to free up space in the context window.

1. Optimizing claw.md

The claw.md file is often the primary source of context bloat.

• Strategy: Keep the file concise. The presenter demonstrated that reducing a file from 910 lines to 33 lines saved 4% of the total context window.
• Proactive Management: Add a specific instruction to your claw.md file: "When the context exceeds 50%, suggest starting a new conversation or using sub-agents." This forces the AI to monitor its own resource usage.

2. Workflow Segmentation (Skills)

Instead of a monolithic claw.md file, break workflows into modular "skills."

• Methodology: Create separate files for distinct tasks (e.g., LinkedIn posts, email replies, proposal generation).
• Benefit: By only invoking the specific skill needed for a task, you avoid loading irrelevant instructions into the context window. This can reduce context usage from 45% down to 27%.

3. Efficient File Referencing

• Reusable Templates: Rather than "baking" long instructions into every skill, create separate reference files (e.g., tone.md, banned_phrases.md). Reference these only when necessary.
• Large File Handling: Do not paste large transcripts directly into the chat. Instead, save them as files in your project directory and instruct Claude to read the file. This reduced context consumption from 71% to 38% in the provided example.

4. System Commands for Context Control

Claude Code provides built-in commands to manage and monitor resources:

• /context: Displays a breakdown of current token usage by category (MCP tools, memory, skills).
• /clear: Resets the conversation entirely.
• /compact: Summarizes the conversation history into a smaller prompt, allowing you to continue the session without hitting the limit. You can specify which information must persist during this process.

5. Managing Memory and MCP Connectors

• Memory Cleanup: Claude stores "memories" about your projects and preferences. Use the prompt: "Please check all my memories that you have about me" to audit this data. If irrelevant projects (e.g., old demo builds) are stored, explicitly command the AI to delete them.
• MCP Audit: Unused MCP connectors (e.g., Slack, Airtable) consume significant tokens. Use cloud MCP list to view active connections and remove unused ones via claude.ai/settings/connectors.

6. Scaling with Sub-agents

For complex tasks, avoid using a single thread.

• Process: Use sub-agents to silo responsibilities. For example, if processing a large file, spawn three sub-agents: one for question extraction, one for action items, and one for decision extraction.
• Result: This distributes the context load across multiple instances, preventing any single thread from becoming overloaded.

Synthesis

The performance degradation of Claude over long sessions is primarily a result of "context bloat." By treating the context window as a finite resource, users can maintain high performance through modularity (skills), external file referencing, and proactive management of memory and MCP tools. The most effective strategy is to shift from a "dump everything in" approach to a "just-in-time" retrieval model, utilizing sub-agents and compaction commands to keep the AI focused and efficient.