用智能体写代码前先约定可验证的完成标准
Rule I'd give anyone coding with agents:
AI 导读
Larsen Cundric 提出用智能体写代码的一条规则：在提出改动需求前，先约定双方如何确认改动生效。他给出的验收方式包括一个当前会失败的测试、一条自己能运行的命令、真实运行结果的日志或截图，没有证据的“完成”会被退回。他认为多数失败的智能体会话都始于模糊的需求和缺失的检查，并向读者征集自己信任智能体“完成”的标准。
正文
Rule I'd give anyone coding with agents:
Before you ask for the change, ask how you'll both know it worked.
> a test that fails right now > a command you can run yourself > a log or screenshot of the real thing > "done" with no proof gets sent back
Most bad agent sessions start with a vague ask and no check.
What's your rule for trusting an agent's "done"?