本文由 AI 分析生成
建立時間： 2026-03-28 來源： https://blog.sshh.io/p/how-i-use-every-claude-code-feature
Summary
Shrivu Shankar’s comprehensive reference on using Claude Code features at scale — from individual hobby use (--dangerously-skip-permissions) to enterprise deployment (monorepo CLAUDE.md consuming several billion tokens/month). Covers CLAUDE.md philosophy, custom slash commands, subagents, hooks, and GitHub Actions integration.
Shrivu Shankar 關於大規模使用 Claude Code 功能的綜合參考——從個人愛好（--dangerously-skip-permissions）到企業部署（monorepo CLAUDE.md，每月消耗數十億 tokens）。涵蓋 CLAUDE.md 哲學、自訂斜線命令、子代理、hooks 和 GitHub Actions 整合。
Key Points
- CLAUDE.md is the “constitution” — most important file for Claude Code effectiveness
- Professional CLAUDE.md at 13KB; only document tools used by 30%+ of engineers; allocate “token budgets” per tool’s documentation
- Anti-patterns: (1) don’t @-mention files directly (bloats context), instead pitch when to read them; (2) don’t use “Never X” without an alternative; (3) don’t document — instead fix the CLI API to be simpler
- “Shoot and forget” philosophy: delegate full tasks, judge by final PR not intermediate steps
- The real race is between Anthropic and OpenAI on agentic tools; tool choice often comes down to “vibes” and system prompt feel
- Key insight: being “too in-the-loop” is a signal you’re not delegating properly; sycophantic responses mean you’re watching too closely
Insights
The “token budget for documentation” concept (treating CLAUDE.md sections like ad space with size limits) is a clever forcing function that prevents the bloat that typically accumulates in these files. The “pitch the agent on why and when to read a file” rather than @-mentioning it directly is a subtle but important optimization — it teaches the model conditional file-reading rather than always loading everything. The observation that Claude Code users often over-index on output style rather than final output quality reflects a broader problem in AI tool evaluation.
Connections
Raw Excerpt
The single most important file in your codebase for using Claude Code effectively is the root CLAUDE.md. This file is the agent’s “constitution,” its primary source of truth for how your specific repository works.