So I built a boilerplate that uses Claude's Sub-agents feature to distribute work across specialized agents. Each agent maintains its own context window for specific tasks—requirements analysis, design docs, TDD implementation, quality checks.

For example, my MCP server project that previously hit the 200K token limit completed successfully, processing 770K tokens across agents.

GitHub: https://github.com/shinpr/ai-coding-project-boilerplate

The approach is basically context engineering—giving each agent only the information it needs, when it needs it. Instead of dumping 40+ sections of rule files into one context, agents get 3-5 relevant sections dynamically selected.

What surprised me most was how effective "meta-cognition" was—an agent that pauses before actions to choose the right context. This small step prevented a lot of rework.

Quick start: `npx github:shinpr/ai-coding-project-boilerplate my-project`

More details and metrics: https://dev.to/shinpr/zero-context-exhaustion-building-produ...