本文由 AI 分析生成
建立時間： 2026-01-03 來源： https://x.com/dotey/status/2007217136176148737
Summary
A Chinese-language walkthrough of Boris Cherny’s (the creator of Claude Code) nine practical tips, the surprise being how “vanilla” his setup is — he barely customizes because it works well out of the box. Tips cover running 5+ parallel instances plus 5-10 web tasks with terminal/web handoff, always using Opus + thinking (slower per-response but fewer corrections), a Git-committed team-shared CLAUDE.md updated weekly with “don’t do this” rules, starting most sessions in Plan mode, slash commands and sub-agents for repetitive work, /permissions over --dangerously-skip-permissions, MCP for Slack/BigQuery/Sentry, and self-verification loops.
這是一篇中文整理，講解 Claude Code 之父 Boris Cherny 的九條實戰技巧，意外之處在於他的配置極其「原裝」——幾乎不做客製，因為開箱即用就很好。技巧涵蓋並行開 5 個以上實例加 5-10 個網頁任務並可終端／網頁交接、所有任務都用 Opus 加 thinking（單次較慢但需糾正的次數遠少）、提交到 Git、團隊共用且每週更新「別這樣做」規則的 CLAUDE.md、多數會話從 Plan 模式開始、用斜線命令與子代理自動化重複工作、以 /permissions 取代 --dangerously-skip-permissions、用 MCP 接 Slack/BigQuery/Sentry，以及自我驗證迴圈。
Key Points
- Parallelism: 5 terminal instances (tabs 1-5) + 5-10 web tasks; handoff via
&and--teleport; start tasks from the phone app. - Model: Opus 4.5 + thinking for everything — slower per response but fewer corrections and more accurate tool calls net out faster.
- CLAUDE.md: single team-shared file in Git, updated weekly; @.claude in PRs adds new rules (“compounding engineering”).
- Plan mode (double Shift+Tab) for most sessions; slash commands (e.g. /commit-push-pr) and sub-agents (code-simplifier, verify-app) for repetition; PostToolUse hook for formatting.
- Tip 9 (most important): give Claude a way to verify its own work — raises output quality 2-3x via feedback loops (bash, tests, browser/simulator).
Insights
The headline insight is “无招胜有招” — mastery here isn’t exotic config but deeply understanding stock features and composing them into a workflow. The Opus-over-Sonnet reasoning reframes “speed” to include human attention/correction cost, not just response latency — a recurring theme with the Context Engineering clippings. The self-verification point (give the agent a feedback loop) is presented as the single highest-ROI investment, echoing the broader vault thesis that orchestration and feedback design beat raw model choice. Several tools (CLAUDE.md, slash commands, hooks, MCP) map one-to-one onto bot_vault’s own skill/hook architecture.
Connections
Raw Excerpt
如果 Claude 能验证自己的工作，最终产出质量能提升 2 到 3 倍。