From “show me your code” to “show me your idea”

从“给我看代码”到“给我看你的想法”

From “show me your code” to “show me your idea” Or: the abdication is signed, and the court is empty 2026-07-13 Second in a six-part series on what building with AI agents did to one developer. This piece is where code loses authority — and the question becomes what can still pin “this is what I meant.” 从“给我看代码”到“给我看你的想法”，或者说：退位书已签，法庭已空。2026年7月13日。这是关于AI智能体开发如何改变一名开发者的六篇系列文章中的第二篇。本文探讨了代码失去权威性的时刻，以及当代码不再能作为“这就是我的本意”的证明时，我们还能依靠什么。

Last week someone turned a laptop toward me. A function on the screen, comments neat, tests green. He pointed at a line and said: this is what I meant. I asked: is that what you meant, or what the model smoothed into place? He froze. I froze. Three seconds opened in the air that neither of us would fill. 上周，有人把笔记本电脑转向我。屏幕上是一个函数，注释整洁，测试通过。他指着其中一行说：这就是我的本意。我问：这是你的本意，还是模型润色后的结果？他愣住了。我也愣住了。空气中出现了三秒钟的空白，我们谁都没有打破。

In those three seconds I counted two thoughts. First: I shouldn’t have asked. Second: who else is left to ask. Fingers on the trackpad edge — the heat was the machine’s, not an idea’s. Two people facing the same green, neither daring to hit Enter. It wasn’t supposed to be like this — I thought. 在那三秒钟里，我脑海中闪过两个念头。第一：我不该问。第二：还能去问谁呢？手指触碰着触控板边缘——那热量来自机器，而非某种思想。两个人面对着同样的绿色测试结果，谁也不敢按下回车键。我心想，本不该是这样的。

There used to be a tyrant: compiler, tests, exit codes. Talk is cheap, show me the code. Pretty talk died where the run failed. Drop the brick; fakes rang. After the ask you could stand across an exit code — a piece of iron in the middle that would not lie. The brick is still there. He pushed it over. The authority is gone. 过去曾有一个暴君：编译器、测试、退出代码。“空谈误国，实干兴邦”（Talk is cheap, show me the code）。漂亮话在程序运行失败时便会破灭。扔出这块“砖头”（代码），伪劣者自会现形。在那之后，你可以站在退出代码的对面——那是一块不会撒谎的铁证。砖头还在那里，但他把它推开了。权威消失了。

Running is no longer the same as “this is what you meant.” The prettier the sample, the more the questioner looks like a nitpicker. And you still can’t drop the question — drop it, and you pretend that running is the same as “this is what you meant.” So some people bring it back: show me your talk. More meetings, spell it out, align the narrative. 程序能运行，不再等同于“这就是你的本意”。示例越漂亮，提问者看起来就越像个吹毛求疵的人。但你依然无法放弃这个问题——如果放弃，就等于假装“运行成功”等同于“这就是你的本意”。于是有些人又回到了原点：给我看你的谈话。更多的会议，更详细的阐述，对齐叙事。

Talk was always cheap. I always took it for empty speech — time burned, energy burned — which is why I would rather stake the final court on code. Code lost its authority. I do not want that road back. So the ask shrinks to: show me your idea. Not inviting talk back in — wanting something that can still be pinned: a tradeoff, a fork, whose fault if wrong. 谈话向来廉价。我一直认为那是空谈——浪费时间，浪费精力——这就是为什么我宁愿把最终的裁决权交给代码。代码失去了权威，我不想回到过去。所以，要求缩减为：给我看你的想法。不是为了让空谈回归，而是为了寻找某种依然可以被锚定的东西：一种权衡，一个分叉点，如果错了，是谁的责任。

Open your mouth, and it becomes talk again. Models say it too, often smoother. The waste I hated most came back in a different shirt. I am still beside him, trackpad heat still on my fingertip, not knowing whether to say “I believe you” or “prove it again.” Both feel wrong. Swallow; throat goes dry. 一旦开口，就又变成了空谈。模型也能说，而且往往说得更顺溜。我最痛恨的浪费，换了件马甲又回来了。我依然站在他身边，指尖还残留着触控板的热度，不知道该说“我相信你”还是“再证明一次”。两者似乎都不对。咽了口唾沫，喉咙发干。

Two days later I asked him again. Softer: not “did you mean this,” but “why write it this way.” He answered fast. Boundaries first, then retry on failure, then why not the other path. Clean. I almost convicted him again for being too smooth. When he finished, I changed the knife — not restatement, a fork: “If the retry budget is cut in half, what breaks first?” 两天后，我再次问他。语气更柔和了：不是问“这是你的本意吗”，而是问“为什么这样写”。他回答得很快。先是边界条件，然后是失败重试，接着解释为什么不选另一条路径。很清晰。我差点又因为他回答得太顺畅而给他定罪。当他说完，我换了把刀——不再是复述，而是抛出一个分叉点：“如果重试预算减半，什么会先崩溃？”

He stopped. Not the pause of swapping connectives. After the stop, he named something that had not appeared in the first speech — the order of cache invalidation. That inch was his. A model can answer a fork too, often more completely. I asked again: if we do it in the order you said, which step blows up first online — which sentence do you own? 他停住了。那不是在斟酌措辞的停顿。停顿之后，他指出了一个在第一次陈述中未曾出现的东西——缓存失效的顺序。那一寸逻辑属于他。模型也能回答分叉问题，而且往往更全面。我再次追问：如果我们按你说的顺序执行，线上哪一步会先炸掉——哪一句话是你自己负责的？

He pointed at the one he had just spoken. That point mattered more than answering completely. I had been measuring wrong: smooth is not evidence of guilt; standing behind the fork is. I asked myself the same way. At night, facing a stretch I had just “written” — I had clicked Accept. The comment announced intent. I wiped the whole intent block, stared at the blank for a long time, then typed. 他指了指刚才说的那句话。那个指点比回答得完整更重要。我之前的衡量标准错了：顺畅不是罪证；敢于为分叉点负责才是。我用同样的方式审视自己。深夜，面对一段我刚刚“写”出的代码——我点击了“接受”。注释声明了意图。我删除了整个意图块，盯着空白看了很久，然后开始打字。

What came out was almost as smooth as what I erased. Smooth enough to sicken me. Deleted again. Typed again. The third pass was still smooth. Finger on Save, I thought: if it writes smooth, does that mean it isn’t mine. No. Reaching the same logic a second time can be called understanding. 写出来的东西几乎和我删掉的一样顺畅。顺畅得让我恶心。再次删除。再次打字。第三遍依然顺畅。手指悬在保存键上，我想：如果它写得顺畅，就意味着不是我的吗？不。第二次推导出同样的逻辑，可以称之为理解。

What I feared was not smooth — it was having no fork I could answer, while still using “I hesitated” to prove it was mine. I added a fork in my head, answered it poorly, and still pressed Save. The cursor moved on. I did not become more myself, or less. 我恐惧的不是顺畅，而是没有我可以回答的分叉点，却还要用“我犹豫过”来证明这是我的作品。我在脑海中增加了一个分叉点，回答得很糟糕，但还是按下了保存。光标继续移动。我并没有变得更像自己，也没有变得更不像。

Later I thought: asking can move a person and still fail to pin a lie, so nail. Nail every claim into something that runs and shows a side effect. One night I nailed eleven. The ninth went red. I fixed it, ran again, red to green. While fixing, my hand was hot — briefly, like heat on loan. When all eleven went green, the terminal lit in a row. 后来我想：提问可以触动一个人，却依然无法戳穿谎言，所以要钉死它。把每一个主张都钉在能运行并产生副作用的东西上。一天晚上，我钉了十一个。第九个变红了。我修复它，再次运行，从红变绿。修复时，我的手是热的——短暂地，像借来的热量。当十一个全部变绿，终端亮起了一排绿灯。

I sat with that row and waited for “this is what I meant” to sound inside. It didn’t. That sit cleared the old books. An exit code only ever proved one thing: it runs. It never proved “this is what you meant.” We treated those as the same passport and called the illusion rigor. The model did not remove the final court — it tore the illusion open. 我坐在那一排绿灯前，等待着“这就是我的本意”在心中响起。但它没有。那次静坐清空了旧账。退出代码只证明了一件事：它能运行。它从未证明“这就是你的本意”。我们把这两者当作同一本护照，并称这种幻觉为严谨。模型并没有拆除最终的法庭——它只是撕开了幻觉。

The green was steady. “This is what I meant” still did not sound. Nails can pin a lie. Nails cannot pin whose idea it was. I did not close his laptop. I did not rewrite that comment for him. The cursor still blinks after “this is what I meant.” I no longer expect it to testify for anyone. 绿色依然稳定。“这就是我的本意”依然没有响起。钉子可以钉住谎言，但钉不住想法属于谁。我没有合上他的笔记本电脑。我没有为他重写那条注释。光标依然在“这就是我的本意”之后闪烁。我不再指望它为任何人作证。

What I can require is only a sentence that can bite me later: which tradeoff I own, and if it is wrong, it is mine. That sentence can live in a PR, or only in the mouth. What matters is not a tidy format — it is whether a person will still stand behind a choice. Those three seconds ended in stopping — stop asking green to prove both “it runs” and “this is what I meant.” 我唯一能要求的是一句日后会反噬我的话：我承担了哪种权衡，如果错了，那就是我的责任。这句话可以存在于PR中，也可以只存在于口头。重要的不是整洁的格式，而是一个人是否依然愿意为自己的选择负责。那三秒钟以停止而告终——停止要求绿色测试同时证明“它能运行”和“这就是我的本意”。