Quartz 5

Boris 的 Claude Code 私人配置技巧

Properties1
tagsnote, ai, claude-code

3 min read

Summary

Boris, the creator of Claude Code, shares his deliberately minimal personal configuration. He runs 5 parallel Claude processes in the terminal plus 5-10 web instances, uses Opus with thinking-for-everything, leans on a team-shared CLAUDE.md, slash commands, sub-agents, hooks, and pre-approved permissions. His single biggest tip: give Claude a reliable way to verify its own work, which improves output quality 2-3x.

Claude Code 的開發者 Boris 分享他刻意保持精簡的個人配置。他在終端並行運行 5 個 Claude 進程，外加 5-10 個 Web 實例，使用 Opus 並開啟全程思考，依賴團隊共享的 CLAUDE.md、斜杠命令、子代理、Hook 與預先核准的權限。他最重要的一條建議：給 Claude 一個可靠的自我驗證機制，能讓產出品質提升 2-3 倍。

Key Points

  • Parallelism over heavy customization: 5 terminal tabs (numbered 1-5 with system notifications) plus 5-10 web/iOS sessions.
  • Opus with thinking-for-everything is faster in practice despite being larger, because it needs less intervention.
  • Team-shared CLAUDE.md committed to git; every Claude mistake gets added so it isn’t repeated. PRs can tag @.claude to append rules (compounding engineering).
  • Most sessions start in plan mode (Shift+Tab twice); a good plan lets auto-accept finish in one pass.
  • Slash commands in .claude/commands/ automate inner-loop workflows (e.g. /commit-push-pr); sub-agents like code-simplifier and verify-app handle recurring PR tasks.
  • PostToolUse hook formats generated code; permissions pre-allowed via /permissions instead of —dangerously-skip-permissions.
  • For long tasks: background verification agents, stop hooks, and the ralph-wiggum plugin in sandboxed environments.

Insights

The recurring theme is that Claude Code works best out of the box and the highest-leverage investments are organizational (shared CLAUDE.md, committed commands/settings) rather than personal tweaks. Verification is framed as the single most important factor — the configuration exists mostly to make Claude’s feedback loops tight and deterministic, not to micro-manage its behavior.

Connections

Raw Excerpt

要想让 Claude Code 发挥最佳效果，最重要的一点可能是给 Claude 提供验证其运行结果的方法。如果 Claude 有这种反馈机制，最终结果的质量将提升 2-3 倍。

Graph View