Okay, now that we have all these tools inside Claude Code, has anybody figured out what would be the best practice? I'm talking about custom slash commands, the agents and also the new skill feature. And how to reduce the Claude.md overhead and also optimize the token windows. I'm thinking about having a lean Claude.md file for acting like an orchestrator and then calling multiple sub-agents, and each sub-agent would have its own skill. So we will have a very long running main task and saving its context window, while delegating all the subtasks to the agents with their specialization.