One Prompt Coding

欢迎回来，继续深入探讨。

Welcome back to the deep dive.

今天是2026年2月17日。

Today is 02/17/2026.

在我们翻开桌上这一堆报告之前，先花几秒钟，回过头想想过去的事。

And before we get into the stack of reports I have on the desk here, I want you to just take a second and mentally rewind the clock.

哦，我挺期待接下来要讲的。

Oh, I like where this is going.

倒回整整一年前。

Go back exactly one year.

想想2025年初的时候，我们处在什么状态。

Think about where we were in, you know, early twenty twenty five.

感觉像隔了几个世纪，对吧？

It feels like a different geological era, doesn't it?

确实如此。

It really does.

我翻了翻去年的笔记，那时候我们还围在一起聊周末搞点‘氛围编程’的事儿。

I was looking through my notes from last year, and we were all sitting around talking about vibe coding on the weekends.

感觉挺实验性的。

It felt experimental.

是啊。

Yeah.

挺可爱的。

It was cute.

确实挺可爱的。

It felt cute.

你让AI写个Python脚本。

You'd get an AI to write a Python script.

结果它就崩了。

It would break.

你修好之后，觉得自己像个魔法师——毕竟你让一个小球在屏幕上弹起来了。

You'd fix it, and you felt like a wizard because you made a ball bounce on the screen.

但看看今天现实的情况，从Reddit上的讨论到行业深度分析，那个时代已经一去不复返了。

But looking at the reality today from Reddit threads to heavy industry analysis, that era is gone.

彻底结束了。

Completely.

我认为可以说，AI编程的业余阶段已经基本结束了。

The hobbyist phase of AI coding is, I think it's fair to say, effectively over.

我们现在看到的是，资深架构师们正在逐步退出传统的编码岗位。

We are seeing a fundamental shift where senior architects are retiring from traditional coding roles.

这并不是因为他们年纪大了。

And it's not because they're getting old.

完全不是。

No, not at all.

而是因为工具已经成熟到足以直接用于生产，软件工程师的职责已经彻底变了样。

It's because the tools have become so production ready that the job description of software engineer has mutated into something else entirely.

而这，正是我们今天深度探讨的主题。

And that is the mission of today's deep dive.

咱们今天要深入剖析这个新的现实。

We're gonna unpack this new reality.

我们会分析三个巨大的转变。

We're gonna look at three massive shifts.

首先是这场单次提示带来的革命，它让我们重新质疑2025年对AI代理的所有构想。

First, this one prompt revolution that is making us question everything we built in 2025 regarding AI agents.

对。

Mhmm.

其次是AI包装类初创公司的市场彻底崩盘。

Second, the absolute cratering of the AI wrapper startup market.

如果你正在做这类项目，一定要认真听这部分内容。

If you're building one, you really need to listen to this.

最后，我们还要面对负责任AI与五角大楼和Anthropic之间一场复杂交织的军事行动之间的激烈冲突。

And finally, we have navigate a very messy collision between responsible AI and a military operation involving the Pentagon and Anthropic.

这典型地反映出技术发展远远快于政策制定，也远远超出了我们调整工作方式的能力。

It's a classic case of technology outpacing policy and frankly outpacing our own ability to adapt our work flows.

对。

Mhmm.

我们看到这些工具的进步速度，快到根本来不及写教程。

We're seeing tools improve faster than we can even write the tutorials for them.

那我们先从技术本身说起，特别是这个‘一个提示’的新现实。

So let's start with the technology itself, specifically the one prompt reality.

我们发现了一个很有趣的案例，来自一位Reddit用户，他说自己用一个提示就构建了一个完整的全栈CR Road应用。

We found this fascinating case study from a Reddit user who claims to have built a completely functional full stack CR Road application.

应用。

Application.

没错。

Right.

对没接触过开发的朋友解释一下，CR Road就是创建、读取、更新、删除，这是任何应用的基础架构。

And for anyone listening who isn't a dev, CR Road just means create, read, update, delete, it's the basic plumbing of any app.

不对。

No.

但这可不是什么简单的‘你好，世界’页面。

But this wasn't just a hello world page.

对。

No.

这里的底层架构出乎意料地复杂。

The plumbing here was surprisingly sophisticated.

是的。

Yeah.

整个技术栈包括 Next.js 作为框架，

We're talking about a stack that includes Next.

UI 组件用的是 ShedCN，现在特别流行，

Js for the framework, ShedCN for the UI components, which is very trendy right now.

特别火，

Very trendy.

数据库用的是 Neon，安全方面用的是 BetterAuth。

Neon for the serverless database and BetterOff for security.

这真是个相当现代的技术栈。

That is a serious modern stack.

这种配置，中级工程师通常得花一周时间来搭建。

That's what you'd expect a mid level engineer to spend a week setting up.

最让人惊讶的是，他们只用了一个提示词就让 GPT-5.3 Codex 完成了所有工作。

And the kicker is, they did it with a single prompt to g p s d 5.3 codex.

就一个提示词。

One prompt.

就一个提示词。

One prompt.

没用什么规划模式。

No planning mode.

后端也没跑一堆复杂的子代理。

No complex sub agents running in the background.

根本没人手把手教。

No hand holding.

他们只是描述了一下应用，模型就一次性生成了完整可运行的代码。

They just described the app and the model spit out the entire thing working on the first try.

这才是真正的转折点。

This is the disruption point.

如果你回看2024年底到整个2025年，行业普遍的看法是得靠智能代理。

If you look back at late twenty twenty four and all of 2025, the industry consensus was agents.

没错。

Right.

人人都在搞代理系统。

Everyone was building agents.

连我们自己也觉得，大语言模型太笨了，没法把整个项目装在脑子里。

Everyone including us said that LLMs were too dumb to hold a whole project in their head.

我们觉得上下文窗口太小，推理能力也太脆弱。

We thought the context window was too small or the reasoning was too fragile.

所以当时觉得，只能靠搭建复杂的链条来解决。

So the solution was to build these complex chains.

一个代理负责规划，一个代理写数据库代码，另一个代理来审查代码。

One agent plans, one agent codes the database, one agent critiques the code.

我们花了数十亿美元来构建这些代理工作流。

We spent billions building these agentic workflows.

本质上就是在电脑里模拟一支开发团队。

Basically trying to simulate a team of human developers inside the computer.

你有项目经理机器人、后端机器人、前端机器人。

You have the project manager bot, the backend bot, the frontend bot.

没错。

Exactly.

但这个GPT 5.3的基准测试表明，模型本身的智能水平——也就是我们说的模型进步速度——正在快过我们为它们开发工具的速度。

But this GPT 5.3 benchmark suggests that the raw intelligence of the models, what we call model velocity, is improving faster than our ability to build tools for them.

模型根本不需要专门的规划代理，因为它的推理窗口和理解能力已经强大到足以一次性把握整个架构上下文。

The model didn't need a planning agent because its inference window and reasoning capabilities have grown so large it could hold the entire architectural context at once.

是的。

Yeah.

它直接就明白了任务要求。

It just understood the assignment.

这可真有点让去年忙活搞代理框架的人尴尬了，对吧？

That's well, it's a little embarrassing for everyone who spent the last year building agent frameworks, isn't it?

感觉我们搭了一整套复杂的脚手架，结果大楼自己凭空冒出来了。

It feels like we built a complex scaffolding for a building and then the building just appeared out of thin air.

某种程度上，确实如此。

In a sense, yes.

社区讨论的共识是，对于中等规模的任务，这些复杂的代理流程可能已经过时了。

The consensus from the community discussion is that for mid sized tasks those complex agentic workflows are arguably becoming obsolete.

模型实际上在说：我不需要你帮我规划这个。

The model is effectively saying, I don't need your help planning this.

我直接就能看到解决方案。

I can just see the solution.

但让我换个角度提个问题。

But let me play devil's advocate here.

Reddit上的演示是一回事。

A Reddit demo is one thing.

但运行银行或医院的生产软件又是另一回事。

Production software that runs a bank or a hospital is another.

你总不能靠一条指令就搞定银行账本吧？

You can't just one prompt a banking ledger, right?

还有合规问题、边界情况、安全审计等等。

There are compliance issues, edge cases, security audits.

你说得对，这正是关键所在。

You are absolutely right and that's the nuance.

一条指令在原型、MVP和内部工具上简直太棒了。

One prompt is incredible for prototypes, MVPs, minimum viable products, and internal tools.

要是你们销售团队需要个仪表盘，一说就行。

If you need a dashboard for your sales team, boom, one prompt.

明白了。

Okay.

但生产环境仍然需要严格的测试和维护。

But production environments still require rigorous testing and maintenance.

不过，这两者之间的差距正在缩小，对你这个听众来说，关键在于抓住这个核心要点。

However, the gap is closing and the strategic takeaway for you, the listener, is crucial here.

别再纠结于具体工具的操作流程了。

Stop obsessing over specific tool workflows.

所以不用花三个月去学什么智能代理X，或者什么完美的思维链流程？

So don't spend three months learning how to use agent x or the perfect chain of thought workflow?

没错。

Correct.

别去学工具本身，要学背后的原理，把精力放在需求分析上，学会怎么写出全面又有效的提示词，学会怎么验证输出结果。

Don't learn the tool, learn the principle, invest in requirements gathering, learn how to structure a massive comprehensive prompt effectively, learn how to test the output.

工具下个月就会变，GPT 5.3到夏天就过时了。

The tools will change next month, GPT 5.3 will be old news by summer.

对。

Right.

但能清晰定义问题的能力。

But the ability to define a problem clearly.

这才是唯一能规模化的能力。

That is the only skill that scales.

这就像学开车，是学开某一款车，还是学怎么应对路况。

It's the difference between learning to drive a specific make of car versus learning how to navigate traffic.

车会换，路却不变。

The car changes, the road remains.

没错。

Precisely.

而且说实话，路况正在变得越来越好，这就引出了Claude。

And honestly the road is getting smoother, which brings us to Claude.

对。

Right.

GPT 5.3在秀肌肉的时候，Claude Code做的却是件让我们久等了的真正提升体验的事。

While GPT 5.3 is flexing raw power, Claude Code is doing something that feels honestly, it feels like a quality of life upgrade we've been desperate for.

他们在测试一个叫自动记忆的新功能。

They are AB testing a new feature called Auto Memory.

这太重要了。

This is huge.

如果你现在用AI写代码，肯定懂这种痛苦。

If you code with AI right now, you know the pain.

每次新开一个聊天会话，都像从零开始。

Every time you start a new chat session, it's a blank slate.

你等于每次都在跟AI第一次见面。

You're essentially meeting the AI for the first time.

每次都得重新说明：我用TypeScript，我偏好这个代码规范，还有昨天我们说好不用那个库了。

You have to re explain, hey, I use TypeScript or I prefer this specific linting rule or remember we decided not to use that library yesterday.

就像每天早上都来一个新实习生，还全忘了之前的事。

It's like having a new intern every single morning who has total amnesia.

每天前二十分钟都得从头给他补课。

You spend the first twenty minutes just catching them up.

这个比喻太贴切了，Auto Memory 正好解决了这个问题。

That is the perfect analogy and Auto Memory fixes that.

它能在不同会话之间持续学习并保留上下文。

It provides persistent context learning across sessions.

它会自动记住你的架构决策、编码风格和常用模式。

It remembers your architectural decisions, your coding style, your common patterns automatically.

那怎么知道自己有没有这个功能呢？

Okay, so how do you know if you have it?

如果你是用户，可以运行 /memory 命令，看看自己是不是在测试名单里。

If you're a user, you can actually check if you're part of the test group by running the slash memory command.

在 Claude Code 里直接输入 memory 就行。

Literally just memory in Claude Code.

对我们这些还没轮上的用户来说就有点倒霉了，我知道他们因为基础设施压力在慢慢 rollout。

And for those of us who unlucky for us don't have the feature yet because I know they are rolling this out slowly due to infrastructure load.

那在没上线之前，有什么办法能临时模拟一下吗？

Is there a way to fake it?

其实有个办法，而且说实话，这本来就是你应该养成的最佳实践。

There is a workaround and honestly it's a best practice that you should probably be doing anyway.

你可以在项目根目录下创建一个叫 clauee 的文件。

You can create a file called clauee.

Md 文件。

Md in the root of your project folder.

明白了，就是个文本文件，里面要写些什么？

Okay, so just a text file and what goes in there?

你可以把它当成一个人工记忆库。

Think of it as a manual memory bank.

把项目的结构、规范、该做什么和不该做什么都写下来。

You write down your project structure, your conventions, your do's and don'ts.

比如明确写上：一定要用箭头函数，或者千万别用这个已经弃用的库。

You literally list things like always use arrow functions or never use this deprecated library.

懂了。

I see.

Claude 在每次会话开始时都会读取这个文件，相当于拥有了持续的记忆。

Claude reads it at the start of every session and effectively simulates having that memory.

但那些已经拥有原生功能的用户反映，性能上有明显差别。

But the users who do have the native feature are reporting a significant performance difference.

再也不用费劲提醒 AI 你是谁了。

The friction of reminding the AI who you are is just gone.

这消除了使用工具时的心理负担。

It's removing the cognitive tax of using the tool.

你坐下来就能直接干活。

You just sit down and work.

但这种便捷性也引出了下一个话题，这个话题有点残酷。

But this ease of use brings us to our next topic and this one is a bit brutal.

如果 AI 真的这么强大，还能记住一切，那些只是给 API 包了个漂亮外壳的初创公司该怎么办？

If the AI is this good and it remembers everything, what happens to all those startups that were just putting a pretty face on top of the API?

说白了，就是一场AI带来的行业洗牌。

The rapper apocalypse.

这本来就是早晚的事，但发生的速度比谁都没想到的还要快。

It was inevitable, but it's happening faster than anyone predicted.

为了让大家清楚，所谓AI包装公司，就是拿GPT-4或者Claude，给它套个漂亮的界面，然后卖成服务，比如专门写营销邮件的AI，或者检查代码的AI。

Just to define terms for everyone, an AI wrapper is a startup that basically takes GPT four or Claude, wraps a nice user interface around it and sells it as a service like AI for writing marketing emails or AI for checking code.

对。

Right.

而源头材料里提到的一个社区观察，可以说是最后一击。

And the source material highlights a specific community observation that puts the nail in the coffin.

社区里有一种叫Claude Code Skills的自定义功能，你可以把它加到自己的Claude实例里。

There are these things called Claude Code Skills custom capabilities you can add to your instance of Claude.

明白了。

Okay.

有位网友分享了一个用于视觉回归测试的技能。

Someone in the community shared a skill for Visual Regression Testing.

给我们详细说说这个是什么。

Unpack that for us.

什么是视觉回归测试？

What is Visual Regression Testing?

这是一种质量保证工作。

It's a quality assurance task.

它会检查你修改代码后，网站会不会出现显示异常。

It checks if a website looks broken after you change the code.

按钮是不是往左移了五像素？

Did the button move five pixels to the left?

颜色是不是有点不对？

Is the color slightly off?

字体是不是错了？

Is the font wrong?

通常你得花钱请公司来做这件事。

And usually you pay a company for that.

一般你每个月花二十到五十美元，让一家Sauce公司自动运行这些测试。

Usually you pay a sauce company 20 or $50 a month to run these tests automatically.

而社区自己就把这个做出来了。

And the community just built it.

他们是为了练手才做的。

They built it as a skill.

就几百行代码，利用了AI的视觉识别能力。

It's just a few 100 lines of code that leverages the AI's vision capabilities.

这直接让那每月20美元的订阅变得多余了。

It effectively killed the need for that $20 a month subscription.

真厉害。

Wow.

这对创业者来说才是最吓人的地方。

And here is the scary part for founders.

这些技术现在都免费分享在GitHub和Discord上。

These skills are being shared freely on GitHub and Discord.

使用这些工具的门槛已经降到了零。

The barrier to entry for these utilities has dropped to zero.

这太吓人了。

That is terrifying.

如果你的商业模式就是：我做一件AI也能做的事，只不过加个按钮，那完蛋了。

If your business model is, I do a thing the AI can do but I put a button on it, dead.

没错。

You are.

如果你的创业项目只是个功能包装，也就是你只是把AI已经能做的事再展示一遍，那你就有麻烦了。

If your startup is just a capability wrapper, meaning you're just exposing a feature the AI can already do, you're in trouble.

要想活下去，你得有独特之处。

To survive, you need differentiation.

比如什么？

Like what?

你需要AI没有的专有数据。

You need proprietary data that the AI doesn't have.

你需要为大企业客户提供复杂的合规功能，而原始模型是做不到的。

You need complex compliance features for enterprise clients that a raw model can't provide.

或者你需要去解决真正棘手的技术难题。

Or you need to be solving genuinely hard technical problems.

你不能再当个中间商了。

You can't just be a middleman anymore.

这让我想起iPhone早期的手电筒应用时代。

It reminds me of the flashlight app era of the iPhone.

还记得2010年吗？

Remember 2010?

那时候人们愿意花99美分买个App来打开LED灯。

People paid 99¢ for an app to turn on the LED light.

结果苹果直接把这功能集成到控制中心了，整个行业一夜之间就没了。

And then Apple just built it into the control center and that entire industry vanished overnight.

没错。

Exactly.

Claude Code正在做同样的事，而跟GitHub Copilot只是你编辑器的一个插件不同，Claude Code是想打造一个平台。

Claude Code is doing the same thing and unlike GitHub Copilot which is really just an extension of your text editor, Claude Code is trying to be a platform.

现在有人在讨论为这些能力建立一个官方市场。

There is talk of an official marketplace for these skills.

这不仅仅是帮你写代码，更是要扩展AI实际能完成的任务范围。

It's not just about helping you write code, it's about expanding what the AI can actually execute.

所以AI不只是写手，更是执行者。

So the AI isn't just a writer, it's a doer.

但我得反驳一下，说这只能用于小脚本或测试工具的说法。

But I want to push back on the idea that this is just for small scripts or testing tools.

我们参考的其中一个资料介绍了Facet项目。

One of the sources we looked at profiled the Facet project.

这可不是什么玩具。

This isn't a toy.

没错。

No.

绝对不是。

Absolutely not.

Facet照片分析引擎是个大工程。

The Facet Photo Analysis Engine is a beast.

这是一个开源的Python工具，代码量有三万七千行。

It's an open source Python tool, about 37,000 lines of code.

三万七千行。

37,000 lines.

这个是用Claude Code开发的吗？

And this was built with Claude Code?

很大程度上靠它帮忙写的。

Heavily assisted by it.

没错。

Yes.

这个工具能用多个计算机视觉模型分析照片库。

The tool analyzes photo libraries using multiple computer vision models.

它会根据美感、构图、清晰度和色彩给图片打分。

It scores images based on aesthetics, composition, sharpness, color.

这玩意儿非常精细。

It's very sophisticated.

那这说明这些工具的极限在哪里？

So what does this tell us about the limit of these tools?

三万七千行代码可不是靠一条指令就能搞定的。

Because 37,000 lines isn't a one prompt situation.

你不可能只说‘帮我做个Facet’，然后就撒手不管了。

You don't just say build me facet and walk away.

这说明我们找到了一个完美的平衡点。

It tells us we hit a sweet spot.

这个项目足够复杂，一个人单独开发得花好几个月，但同时它的目标又足够明确，让AI能有效搭建框架。

This project is complex enough that a human would take months to build it alone, but it's well defined enough that the AI can structure it effectively.

所以这是关于杠杆效应？

So it's about leverage?

完全就是杠杆效应。

It's all about leverage.

开发者利用了现有的库、视觉模型和网页框架。

The developer leveraged existing libraries, vision models, web frameworks.

没有重新造轮子，结果原本需要几个月才能完成的项目，被压缩到了几周，甚至可能只有几天的主动工作时间。

Instead of reinventing the wheel, the result is that a multi month project was compressed into weeks, maybe even days, of active work.

这种时间上的压缩，正是这里反复出现的主题。

That compression of time is the recurring theme here.

我们正在缩短从想法到实现之间的时间。

We are compressing the time between idea and execution.

但正如我们所知，当加速得这么快时，迟早会撞上一堵墙。

But as we know, when you accelerate power that quickly, you eventually run into a wall.

有时候这堵墙是伦理问题，有时候是地缘政治因素。

And sometimes that wall is ethical, sometimes it's geopolitical.

这就引出了五角大楼的话题。

And that brings us to the Pentagon.

是的。

Yeah.

这件事正在发生，情况很复杂。

This is a story that's developing right now, and it's tricky.

我们得说得准确一点。

We need to be precise here.

有报道称，美军在抓捕委内瑞拉的尼古拉斯·马杜罗的行动中使用了Anthropic公司的ClaudeAI。

Reports are coming out that the US military used Claude Anthropic's AI during an operation to capture Venezuela's Nicolas Maduro.

这可是个重大争议点。

This is a major flashpoint.

我们得严格依据报道的事实，但核心矛盾确实引人深思。

We have to be careful to stick to the facts as reported, but the core conflict is fascinating.

Anthropic一直把自己定位为负责任的AI公司。

Anthropic brands itself as the responsible AI company.

没错，这正是他们的核心主张。

Right, that's their whole thing.

他们的整个宪法式AI框架，都是围绕安全和伦理准则建立的。

Their entire constitutional AI framework is built around safety and ethical guidelines.

他们对军事用途有严格限制，特别是涉及直接行动的场景。

They have strict usage policies against military applications involving direct farm.

所以当他们发现自己的软件可能被用在抓捕行动中时。

So when they found out their software might have been used in a capture operation.

据称，他们主动联系了国防部，问：我们的软件是不是被用了？

They reportedly reached out to the Department of Defense to ask, hey was our software used?

如果是，具体是怎么用的？

And if so, how?

报告说，五角大楼对这个问题非常不高兴。

And the report states that the Pentagon did not take kindly to that inquiry.

他们不喜欢被人这么问。

They didn't like being asked.

对。

No.

据报道，国防部甚至威胁要惩罚任何提出这种质疑的人。

The DOD is reportedly threatening punishment for even the questioning.

这引发了一个关于软件本质的极大问题。

This raises such a massive question about the nature of software.

如果我造了一把锤子，卖给家得宝，我不会去问你用它来盖房子还是砸窗户。

If I build hammer and I sell it at Home Depot, I don't get to ask if you used it to build a house or break a window.

卖出去之后，东西就是你的了。

Once I sell it, it's yours.

对。

Right.

但软件就不一样了，尤其是像这种会直接冲击他们服务器的工具，感觉完全不一样。

But software, especially saws, where you're hitting their servers, feels different.

确实不一样，因为你买的不是一件产品，而是一种服务。

It is different because it's not a product you own, it's a service you access.

这里的矛盾在于：开发者该不该控制别人怎么用你的工具？Anthropic 就想确保他们的 AI 不会被用来做伤害性的事，按他们自己的标准。

The tension here is, should developers have control over the end use of your tools, Anthropic wants to ensure their AI isn't used for harm according to their definition.

还有五角大楼。

And the Pentagon.

五角大楼认为，国家安全行动不属于软件供应商该过问或控制的范围。

The Pentagon argues that national security operations are outside the purview of a software vendor's curiosity or control.

这是两种世界观的碰撞。

It's a collision of world views.

硅谷的理想主义 vs 地缘政治的现实。

Silicon Valley idealism versus geopolitical reality.

没错。

Exactly.

我们在这里并不站队。

And we aren't taking sides here.

这两种观点在各自的体系里都有其内在的逻辑。

Both perspectives have logical consistency within their own frameworks.

但这凸显了‘负责任的AI’绝不仅仅是个营销口号。

But it highlights that Responsible AI isn't just a marketing slogan.

它其实是一片法律和运营上的雷区。

It's a legal and operational minefield.

如果你是用这些接口的开发者，得明白服务条款不再只是走形式的套话，而是实实在在的地缘政治红线。

If you are a developer using these APIs, you need to understand that the terms of service aren't just boilerplate anymore, they're active geopolitical lines.

说到营销，资料里还提到有人批评 Anthropic 的营销方式，让人觉得 Claude 可能有意识？

And speaking of marketing, there's also been some criticism mentioned in the sources regarding Anthropic's marketing suggesting Claude might be conscious?

哦对，就是那种‘我还活着’的感觉。

Oh right, the I am alive vibe.

在这种语境下，这样宣传有帮助吗？

Is that helpful in this context?

大多数行业专家都说不会，这样做只会让情况变得更复杂。

Most industry experts say no, it just muddies the water.

我们面对的是强大的统计工具，也就是下一个词预测器，而不是有意识的生命体。

We are dealing with powerful statistical tools, NEXT token predictors, not sentient beings.

当

When

你把这种强调意识的营销手段和军事应用混在一起时，公众的看法就会变得非常不稳定。

you mix consciousness marketing with military applications, you get a very volatile public perception cocktail.

这制造了本不该有的恐惧，而不是促进技术理解。

It creates fear where there should be technical understanding.

让人一想到的是天网，而不是一个高级的物流工具。

It makes people think Skynet instead of advanced logistics tool.

信息量确实很大，但我想把这次深度探讨拉回现实，给那些正坐在键盘前的听众一些实际的建议。

It's a lot to process, but I wanna ground this deep dive back into something practical for the listeners who are actually sitting at their keyboards right now.

我们聊了宏观的变革、说唱歌手的商业崩塌、战争伦理，现在咱们聊聊打字吧。

We've talked about the high level revolution, the business collapse of rappers, and the ethics of war, But let's talk about typing.

我们在社区讨论里找到了一些非常具体的高手技巧，教你怎么真正适应这套新工作流。

We found some really specific pro tips in the community discussions on how to actually manage this new workflow.

接下来是实战技巧部分。

This is the practical mastery section.

如果你真想用好这些工具，开发者眼下最大的问题就是上下文过载。

If you want to use these tools effectively, the biggest problem developers are facing right now is context dumping.

就是把整个代码库全复制粘贴进聊天框，然后听天由命？

Is that where you just copy paste your entire code base into the chat and hope for the best?

这是我写的 spaghetti 代码，帮忙修一下。

Here's my spaghetti code, please fix it.

没错，你把整个仓库一股脑贴上去，结果 AI 就懵了。

Exactly, you dump the whole repository and the result is the AI gets confused.

它会凭空捏造根本不存在的依赖，因为被太多噪音淹没了。

It hallucinates dependencies that don't exist because it's drowning in noise.

这就好比你只想让工人修个同步问题，却把整座城市的施工图全扔给他。

It's like handing a contractor the blueprints for the entire city when you just want them to fix a sync.

那该怎么解决呢？

So what's the fix?

我们该怎么筛选出有用的上下文？

How do we curate that context?

有选择性地提供上下文。

Selective context.

有些工具比如 repoprint.com 会把你的项目打包成一个文本文件，小项目还行，但官方建议别太依赖。

There are tools like repoprint.com that flatten your repo into a text file, Which is okay for small projects, but the sources advise against overuse.

更好的方法是，如果你在终端用 Claude Code，就用 @文件名 这种方式引用。

The better technique is to use the at filename mention system if you're using Claude Code in the terminal.

明白了。

Okay.

或者创建一个临时的 context.md 文件。

Or create a temporary context dot m d file.

也就是说，只给它完成这个任务真正需要的内容。

So you just give it exactly what it needs for this specific task.

对。

Right.

接下来是社区里公认最有效的提示模板。

But here is the golden prompt that the community swears by.

如果你除了这个视频只记一件事，那就记这个。

If you take one thing away from this episode, take this.

你告诉 AI：如果需要引用没显示的文件里的内容，先问清楚，别自己猜。

You tell the AI, if you need imports from a file not shown, ASK before assuming.

太喜欢这个了。

Oh, love that.

这迫使AI不再乱猜。

It forces the AI to stop guessing.

它让AI明确问题的边界。

It forces the AI to define the scope of the problem.

与其凭空捏造一个不存在的函数，它会停下来说：嘿，我看到你在调用 utils。

Instead of hallucinating a function that doesn't exist, it stops and says, Hey, I see you're calling utils.

但我看不到这个文件。

Js, but I can't see that file.

你能给我看一下吗？

Can you show it to me?

就这一句话，能省下好几个小时的调试时间。

That one sentence saves hours of debugging time.

这简直是金玉良言。

That is gold.

另外，对于那些经常在外奔波的人，我看到有人用iPhone写代码的配置。

Also, for the road warriors out there, I saw a setup for coding on an iPhone.

这听起来简直太赛博朋克了。

This sounded extremely cyberpunk.

确实如此。

It is.

移动端的Claude Code配置用到了TMUX、Termius和Tailscale。

The mobile Claude Code setup involves using TMUX, Termius, and Tailscale.

简单来说，就是用手机远程连接到家里的电脑。

Basically, you SSH into your home machine from your phone.

没错。

Right.

因为Claude Code是个命令行工具，它在终端里运行。

Because Claude Code is a CLI tool, it runs in the terminal.

所以你可以用Tailscale搭建一个安全的私有网络，用Termius在iPhone上当终端，再靠TMUX保持会话不中断。

So you can use Tailscale to create a secure private network, Termius as your terminal on the iPhone, and TMUX to keep the session alive.

所以你坐在咖啡馆里，或者在火车上，对着手机打字，实际上却在操控家里的服务器上那个庞大的AI开发环境。

So you're sitting in a coffee shop or on a train typing into your phone, but you're controlling a massive AI development environment on your home server.

没错。

Exactly.

这说明，你随时随地都能运行这种高级别的开发流程。

It proves that you can run these architect level workflows from anywhere.

你不需要三屏显示器来指挥AI，只需要连上那个‘大脑’就行。

You don't need a triple monitor setup to direct the AI, you just need a connection to the brain.

还有对预算敏感的朋友。

And for the budget conscious.

说实话，如果你频繁迭代，Claude和GPT 5.3的API费用真的会蹭蹭往上涨。

Because let's be honest, the API costs for Claude and GPT 5.3 can rack up if you're iterating a lot.

现在也有一股开源的潮流在兴起。

There is an open source movement happening too.

你可以用一个叫Free Claude Code的工具。

You can use something called Free Claude Code.

它是个代理工具，让你能用 Claude 的命令行界面——就是我们都喜欢的那个简洁终端界面，但把背后的模型换成运行在 Nvidia Nim 上的开源模型，比如 GLM5。

It's a proxy that lets you use the Claude CLI interface, the nice terminal UI we all like, but swap out the brain for open models like GLM5 running on Nvidia Nim.

所以你得到了好用的界面，但用的是更便宜的模型。

So you got the nice interface, but a cheaper brain.

没错。

Exactly.

这样就能享受同样的工作流程，却不用付订阅费，当然了，模型的智能水平会有所不同。

It's a way to get the workflow without the subscription cost though obviously the intelligence level varies.

GLM5 不错，但没法跟 GPT 5.3 比，不过用来学习完全够用了。

GLM5 is good, but it's not GPT 5.3, but for learning, it's perfect.

太厉害了。

Incredible.

我们从那个让 Reddit 震动的 One Prompt 应用，聊到五角大楼和 Anthropic 的对峙，再到那个 rapper 创业公司的倒闭。

We've covered everything from the One Prompt app that shook Reddit to the Pentagon versus Anthropic showdown to the death of the rapper startup.

这期内容信息量真大。

It's been a dense update.

这行当每周都在变。

The landscape is changing weekly.

那咱们就先到这里吧。

So let's wrap this up.

如果我们汇总一下这些信息，就能看出开发者体验已经发生了巨大变化。

If we look at the summary of all these sources, the developer experience has shifted massively.

入门的门槛已经低多了。

The barrier to entry has dropped.

个人项目现在几天就能搞定，不用再花几个月。

Side projects take days, not months.

但对什么是‘架构’的期待却大大提高。

But the ceiling for what is considered architecture has risen significantly.

这才是关键。

That is the key.

如果非让我留给听众一个值得深思的问题，那就是：如果AI能用一条指令就写出完整应用，还能自动记住你的架构风格和偏好，那你真正值钱的就不再是写代码了。

And if I can leave the listener with one provocative thought to mull over today it's this, if the AI can build a full app with one prompt and if it can remember your entire architectural style and preferences automatically the value you bring is no longer writing code.

没错。

Right.

写代码语法已经不再是你的工作了。

Typing syntax is no longer the job.

你现在的工作纯粹是定义问题。

The job is purely defining the problem.

问题是，你准备好不再当一个搬砖的工人，转而成为设计大教堂的建筑师了吗？

The question is are you ready to stop being a builder who lays bricks and start being an architect who designs cathedrals?

因为砖块现在自己在垒了。

Cause the bricks are laying themselves now.

这真是个既令人害怕又令人兴奋的想法。

That is a scary and exciting thought.

别再忙着建造了，开始去定义吧。

Stop building, start defining.

感谢你和我们一起深入探讨。

Thanks for diving in with us.