本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/MushtaqBilalPhD/status/2053829787219595725
Summary
The second tutorial in a series targeting academic researchers, building on 101 with advanced Claude Code patterns: hierarchical CLAUDE.md files (global + per-subfolder), Plan Mode for complex multi-step tasks, Custom Slash Commands for repetitive workflows, Subagents for parallel isolated processing (avoiding “context clutter”), MCP Connectors for external tool integration (Zoom, Google Drive, PubMed), and Hooks/Scheduled Tasks for automated backups and literature scans.
學術研究者 Claude Code 第二部曲，涵蓋階層式 CLAUDE.md、計劃模式、自訂斜線命令、子代理並行處理（避免「上下文污染」）、MCP 連接器，以及自動備份和文獻掃描的 Hooks/排程任務。
Key Points
- Nested CLAUDE.md: global file gives big picture; local files in subfolders (Literature, Chapters, Data) give task-specific rules — Claude reads both
- Plan Mode (Ctrl+Shift+M): Claude writes step-by-step plan before executing; essential for tasks with 3+ steps or multi-folder operations
- Subagents vs Slash Commands: Slash Commands = shortcut (no context window); Subagents = isolated specialist with own context window, doesn’t read main CLAUDE.md
- Subagent use case: parallel critiques — Methodology Auditor + Reviewer 2 run simultaneously, each with clean context, main session uncontaminated
- MCP Connectors listed: Zoom, Google Drive, PubMed, arxiv — connect through Customize → Connect Your Apps
- Hooks auto-trigger on events (e.g., pre-edit backup hook before any Claude edit); Scheduled Tasks run on cron schedule
- Critical rule: never create Hooks/Scheduled Tasks for file deletion, and only automate tasks you’ve done manually at least 4 times
Insights
The “context clutter” concept is the key practical insight: when a single session processes 20 PDFs and then tries to draft a chapter, the accumulated literature context degrades chapter-writing performance. Subagents solve this by providing clean, task-scoped context windows. The parallel critique pattern (multiple subagents with different reviewer personas) is a sophisticated use of Claude’s own uncertainty — rather than asking one Claude to critique from multiple angles (which produces internally consistent but potentially biased output), parallel independent sessions produce genuinely orthogonal feedback.
Connections
Raw Excerpt
Do not create a subagent for minor tasks. Do not give your subagents overlapping responsibilities. Never let your subagent edit your drafts.