Initial message consumes ~24k tokens from injected cache (system reminders, MCP tools, skills)

Description

The very first message in a conversation consumes approximately 24,000 tokens of context before any user interaction, due to cache-injected content including:

• System reminders (multiple blocks)
• MCP server tool definitions (hundreds of deferred tools listed)
• MCP server instructions (e.g., Figma Console workflow instructions)
• Skill listings (30+ skills enumerated)
• CLAUDE.md contents
• Git status snapshot
• Memory system instructions

Impact

• Significant portion of the context window is consumed before the conversation even begins
• Users with multiple MCP servers configured are hit hardest, as each server adds tool definitions and instructions to the injected payload
• Reduces the effective usable context for actual work
• On smaller context windows this is especially painful

Expected behavior

The initial context overhead should be minimized. Possible approaches:

• Lazy-load MCP tool definitions only when relevant (rather than listing all deferred tools upfront)
• Compress or deduplicate system reminder blocks
• Only inject MCP server instructions when those tools are actually being used
• Make the injected content more token-efficient

Environment

• Claude Code CLI
• Multiple MCP servers configured (GitHub, Figma, Jira/Atlassian, JetBrains IDEA, etc.)
• Multiple skills installed