本文由 AI 分析生成
建立時間： 2026-05-07 來源： https://x.com/MushtaqBilalPhD/status/2053829787219595725
Summary
The follow-up to Claude Code 101, covering advanced patterns for long academic projects: hierarchical folder structure with nested CLAUDE.md files, Plan Mode for complex tasks, Custom Slash Commands for repetitive tasks, subagents for parallel/independent research tasks, MCP Connectors for integrating external apps (Zoom, Google Drive, PubMed), and Hooks/Scheduled Tasks for automated workflows. The article addresses context clutter as the core limitation of single-session work.
101 的進階篇，涵蓋長期學術專案的組織策略：巢狀 CLAUDE.md 層級、Plan Mode（複雜任務）、Custom Slash Commands（重複任務）、子智能體（平行獨立任務）、MCP Connectors（外部應用整合）與 Hooks/排程任務（自動化工作流）。核心問題：情境污染（context clutter）是單一 session 的根本限制。
Key Points
- Nested CLAUDE.md: global file in main folder (project constitution) + local files in each subfolder (task-specific instructions); Claude reads both when working in a subfolder
- Plan Mode: Claude writes step-by-step plan before acting; use for tasks with 3+ steps, multiple subfolders, or long outputs
- Custom Slash Commands: markdown files in .claude/commands folder; for repetitive tasks only — keep one command per one specific task
- Subagents: have their own context window and do NOT read the global CLAUDE.md; ideal for parallel critiques, citation checking, methodology auditing
- MCP Connectors: Zoom, Google Drive, PubMed available; connect only project-relevant apps; avoid connecting apps with confidential data
- Hooks: fire automatically on events (e.g., pre-edit safety hook creates backup before editing any chapter); never set up hooks that delete files
- Key rule: do not create subagents for minor tasks; do not give overlapping responsibilities; subagents should produce reports as separate files, never edit drafts
Insights
The context clutter problem described for subagents maps directly to why parallel sessions matter — each agent maintains its own working memory so intermediate reasoning stays isolated. The Reviewer 2 subagent example (critiquing as a hostile reviewer) is the most immediately actionable idea: running genuinely adversarial critique in parallel with constructive critique, without one contaminating the other, is something humans cannot do but subagents can. The prohibition on subagents editing drafts (produce reports as separate files only) is a sound safety principle that mirrors code review best practices.
Connections
Raw Excerpt
“Think of a subagent as a specialist version of Claude Code with its own instructions and, more importantly, its own context window… Unlike a Slash Command, a subagent has one.”