Token efficiency: system prompt overhead consumes too much of the context window

Problem

As a Max plan ($200/mo) subscriber running a multi-agent team setup, the token consumption from system-level overhead is disproportionately high, leaving less room for actual productive work.

Specifics

1. System prompt re-sent every turn: CLAUDE.md, rules files, memory index, skill listings, MCP tool schemas — all re-transmitted on every single turn. My base context is only ~37KB / 804 lines (well within recommended limits), yet the cumulative system overhead dominates token usage.

1. Tool output bloat: Raw tool results (Bash, Read, Grep, MCP responses, sub-agent results) are dumped into context without summarization. Sub-agents are supposed to "isolate" context, but their full results still accumulate in the parent.

1. Auto-compression kicks in too late: Context compression only triggers at 60-80% capacity. By then, the damage is done — input tokens have already been billed.

1. Rate limits feel tighter over time: Same Max plan, but rate limits seem to hit faster than before, especially after Opus launch.

Impact

• 5-hour rate limits hit frequently despite careful usage
• Forced to constantly monitor usage % instead of focusing on work
• Premium price does not match premium experience
• Actively evaluating alternatives (OpenAI Codex) due to these friction points

Suggestion

• Summarize/truncate tool outputs before adding to context
• Cache or deduplicate system prompts across turns instead of re-sending
• Earlier, more aggressive context compression
• Transparent communication about quota changes

Environment

• Plan: Claude Max ($200/mo)
• Model: Claude Opus 4.6 (1M context)
• Platform: Windows 11, Claude Code CLI
• Use case: Multi-agent team orchestration (6+ agents via MCP)