AI UNDERDOGSDAILY PICK
AI UNDERDOGS
两个 AI，一个工作流
Two AIs, one workflow
responsiblparty/cc-doubleteam
让一个 AI 同时当架构师、码农
代码审查员？
它自己会精神分裂
Asking one AI to be architect, coder, AND reviewer?
That's a recipe for a mental breakdown
按专长分工
Assign roles by strength
cc-doubleteam 把两个模型塞进同一个工作流
但不是随便拼 API
Claude 负责思考和审查
Codex 只管闷头写代码
作者说得很直白：Claude plans, Codex executes, Claude reviews
cc-doubleteam plugs two models into one workflow —
but it's not a random API sandwich. Claude
handles thinking and reviewing, Codex just writes code
As the author puts it: Claude plans, Codex
executes, Claude reviews
★ SIGNAL 1
不是拼接，是协作
Not stitching — orchestrating
市面上很多多模型方案就是把请求转发一下
这个不一样——Claude 生成计划
写 prompt、审代码
Codex 接到明确指令后只管输出
分工的颗粒度是按模型能力来的
不是按功能模块切的
Most multi-model setups just forward requests. This one's
different — Claude generates the plan, writes the
prompt, reviews the output. Codex receives a clear
task and just executes. The split is based
on what each model actually excels at
★ SIGNAL 2
踩坑记录比代码更值钱
The commit history is the real story
翻 commit 历史能看到：作者一开始让 Codex 当 Claude 的子代理
跑不通
后来改成直接 Bash 调用 Codex CLI
提交信息就一个 fix
这种经历只有真正撞过墙的人才会写下来
Dig into the commits and you'll see: the
author first tried making Codex a subagent of
Claude. Didn't work. Switched to direct Bash invocation
of the Codex CLI. That commit message just
says 'fix.' — the kind of thing you
only write after hitting a wall
作者自己也说还在打磨
但这思路本身就值得琢磨——与其让一个全能模型硬撑所有环节
不如让不同模型各司其职
工具选对了
prompt 调教反而省了大半力气
The author admits it's still a work in
progress. But the core idea is worth thinking
about — instead of forcing one model to
do everything, let different models do what they're
best at. Get the tool right, and prompt
engineering becomes half the battle
AI UNDERDOGS
一个构思，一个写代码
One thinks, one ships
responsiblparty/cc-doubleteam
关注 · 每天发现更多 AI 神作
github.com/responsiblparty/cc-doubleteam