本文由 AI 分析生成
建立時間： 2025-12-30 來源： https://x.com/kevinma_dev_zh/status/2006044914627535187
Summary
The author describes a personal development workflow that treats Claude Code, Codex CLI, and Gemini CLI as a three-member AI team. Gemini handles planning and design docs, Claude Code reviews the plan then executes, and Codex reviews and optimizes the code — mirroring human team cross-review to catch the bugs Claude introduces on larger tasks. All three agents live in one terminal managed by tmux, with context shared via copy-paste and, more often, written documents.
作者分享將 Claude Code、Codex CLI 與 Gemini CLI 視為三人 AI 團隊的開發工作流。Gemini 負責規劃與設計文件，Claude Code 先 review 計畫再執行，Codex 則 review 並優化程式碼，仿照人類團隊交叉評審來捕捉 Claude 在大型任務上常引入的 bug。三個 agent 都放在同一個 tmux 管理的終端裡，上下文靠複製貼上、更多時候靠寫文件來同步。
Key Points
- Claude Code is fast but reliably introduces bugs on large tasks and can loop when self-fixing.
- Workflow: Gemini plans → Claude Code reviews plan and executes → Codex reviews and optimizes code.
- Cross-review between models mirrors human team review.
- Gemini chosen for planning purely for its writing style; Codex output is tiring, Claude’s docs verbose.
- UI tweaks skip the pipeline — handled directly by Claude Code.
- tmux manages all three agents plus nvim and a service window; context synced via docs.
Insights
The key non-obvious move is using a different model to review, exploiting the fact that an LLM will sycophantically agree with bugs in its own code (“对对对,确实有这个问题”) but a fresh model with no ego in the output catches them. The human’s role collapses to gatekeeper: approving the design and making the final call, while models discuss and review among themselves. Document-based context sharing is the pragmatic glue when there’s no shared memory across agents.
Connections
Raw Excerpt
我一般是让 Gemini 做 planning，输出设计文档；Claude Code review 这个 plan，然后负责执行；Codex 依据 planning doc，review 代码并优化。