前言
本篇文章，只是记录一下我最近半年多年，使用AI工具进行日常工作的感悟。
到了如今的这个时代，基本上不使用AI工具很难继续进行工作了。
Claude Code终端
其实我最早的使用的AI工具是Trae，因为这个公司有合作，奈何Token给的太少，根本无法胜任常规任务，于是就在今年年初使用了Claude Code终端，当时使用的模型是GLM4.7（白嫖了同事的模型）以及自己付费买的MiniMax套餐。
有关这里做了哪些工作，其实在上一篇我已经有很多说明了这里就不过多的描述了，有兴趣的大家可以看看。 聊聊我最近都干了些什么，AI 时代的手动撸码人
感悟就是AI确实可以提高效率，因为实战的项目都是代码升级、库更新、技术栈更新，所以很难以点概面，我当时理解，对于特别的复杂的工作，AI还是不太行了。
后来因为各种原因，以及Claude Code自己作死，要求不允许使用Claude Code，于是换工具势在必行。
Codex App桌面端
后来就用Codex App，其实一开始我也是特别抵触的，我好不容易用习惯了Claude Code终端，现在说换就换，我重新培养使用习惯难道就不是成本吗？
于是在B站刷了几个教学视频，实际操作了一番，发现，有Codex App桌面端，还需要啥子自行车。
Codex App桌面端真香！
用上了Codex之后，不论是上传图片、文件、文件夹等其他的都比终端友好很多。
除了编程，使用Codex之后，它可以更加系统的读取整个工程项目了，页面UI的易用性让你忍不住使用它。
同时我除了接触GLM、MiniMax之外，也接触了更多的模型，Deepseek差不多也就是在此时进行了解的。
Deepseek给我的感觉就是夯，但是当时手上有Code Plan，深度使用谈不上。也不太懂模型和模型之间差距怎么会这么大。
真正的转折出现在AI的工作流使用。
AI工作流
之前编码工作，基本上就是在现有代码做小修小补，或者说小迭代，使用AI输入与输出即可，真正见真章的是从0开始的项目，如果需要让AI从0开始工作，AI的工作流必不可少。
GSD
公司大佬当时在内部推荐使用GSD流程，我也在工作中尝试使用，说来这个GSD也挺有意思的，因为原作者某些原因提桶跑路之后，后面只好另起炉灶。
使用GSD发现的模型之间的差异
在使用GSD的过程中，我觉得这个流程还是很不错的，就是Skill太多，有的时候不知道怎么去选择，另外我的Code Plan是MiniMax，在使用过程中，我越来越发现它是真的"蠢"，它不会自己压缩上下文，有的时候Skill明显它自己要触发，它太依旧头铁的自己干，不用Skill，同时会活没干完就不动。
于是没有办法，借用了同事的GLM，自己充值了Deepseek试了一下，发现原来自己写的提示词没问题，纯粹是模型问题导致的问题。
于是MiniMax的Code Plan打水漂了，为了后续开发工作，更换模型势在必行。
Skills for Real Engineers
随着项目的持续进行，以及模型的进化，工作中，我们发现使用GSD的问题会越来越多，并不是GSD不好，而是不适合我们。于是在几次试水之后，我们的AI工作流，迁移到了前端大神matt pocock的AI工作流上面。
使用Skills for Real Engineers的心得
首先，你要使用这个Skill，至少保证模型能力。这个阶段我自己充值了Codex Plus的Code Plan，另外配合Deepseek，简直一发不可收拾。
另外随着Grill me/Grill with Docs → toSpec → toTickets → Implement的主流程走，最难的是第一步Grill。
经常一个需求输入，Grill一来啪啪啪30个问题很让人崩溃，幸好后面这个Skill升级了，有推荐回答，但是，但是，但是（重要的事情说三遍），作为开发者，你手上握着最终决定权，推荐的回答也许不是你期望的需求，所以一定要认真审核！
后面的流程就很简单了，toSpec → toTickets → Implement，AI进行Codereview，自己验收一下成品。
除了有些费token，我认为这一套走下来还是经得起推敲的。
AI，我，以及感悟
上面的流水账也是只是记录我的使用体验，下面这些是我自己的一些感悟。
有AI了，上班还是很累
其实单纯写代码不累，因为困难的代码基本不会写了，技术上的瓶颈基本不存在了，但是随着AI越来越会写代码，我自己职责边界却变了，这些反而让人更辛苦。
比如之前需求输入到我这里，我理解，然后写代码，自测即可，但是现在变成了，我的理解，如何变成输入给AI，让AI写出的预期功能代码，绞尽脑汁写出提示词成为了一个比较麻烦的问题，如果不是有AI流程，toSpec → toTickets，我觉得写一份规范的输入是一件废脑力的事情。
另外，就算不写代码了，我会不得不审核大量的文档，这一点也不轻松，甚至在项目初期，我也会在一个功能模块下面，去审核AI写的代码是否符合预期，虽然有各种md文件与规则去约束AI写代码，但是还是按流程得过一遍。
另外，从领导的角度看，我手上一定会有AI，工作量与排期都比之前加快了很多。而且公司对AI的token支持不太友好，目前真的就是付费上班，既然付费上班，我就期望自己按计划写完功能，准时下班，但是并不是那么回事。
有AI了，心态变了很多
之前，我总觉得AI就算发展的很快，想要干掉移动端开发还是有点难度的，因为移动端开发有别于前端，它非常依赖多模态的模型（UI需要），同时Xcode的配置又非常复杂，但是随着AI的发展，甚至Xcode27的配置改为了json，其实都是在为AI服务。
我从一开始有点拒绝到接受，和积极拥抱，转变非常之快。
目前唯一的焦虑是，我到底可以干什么？我在自己的岗位上面是不是可有可无？我会不会被干掉？
有AI了，这几个月都干了些什么？
之前的RxStudy还只是我用自己的项目去做迁移，随着公司项目的接入。
我在7~9月份，在一个项目完成了React Native在双端的落地；同时在另外一个项目上，完成了KMP以及CMP在双端的落地，这个项目期望也有HarmonyOS Next端，我们甚至做完了一部分Quickly+KMP在HarmonyOS Next端的验证，以及完整的鸿蒙Flutter的开发。
作为一个iOS开发，很多技术栈我都不太了解，甚至不了解，但是最后还是落地，项目的配置我不用关心，我只有看看代码写的看不看得懂即可。
现在大家都调侃AI叫天才程序员，目前我除了可以看看它写的代码结构与分层，大部分时间我都很难再阅读它写的代码了，因为它的写法比我要高级，大部分时候它的架构与思路比我优秀。
有了AI，写代码已经没有乐趣
这绝对是真实的感受，当人类停止探索的时候，已经没有乐趣，这个就和打游戏一样，当你面对未知广袤的地图，一步步探索一步步推进还会获得乐趣，目前AI写代码没有什么技术难点，在我看来，只要Token管够，时间周期允许，合理的需求都可以开发出来。
而我们，就像流水线上面的工作人员，一步步去检查与核对。
乐趣变成了如何有个好的点子，甚至是一个有趣的交互或者动画，从繁重而重复的CV工作下来之后，我会更多的参与需求的输入，需求是否合理，能不能跑完。
AI一方面，模型强才是关键
我也看过很多人说同一个模型在不同的IDE加持下的表现会有些许差异，但是从我用过的GLM、MiniMax、Deepseek和ChatGPT来看，如果模型本身足够强大，可以做到力大砖飞的程度。
我会经常使用Skill、MCP以及插件，甚至专门去学习如何安装与使用，但是在我们的开发过程中发现，对于一个足够强的模型，说清楚要做的事情，放开权限，那么其他的事情交给它自己就可以了。
同时这里也需要注意，权限的赋予是一柄双刃剑，有便利的同时也有隐患。
好了以上就是我的一些感悟，欢迎大家交流，如果大家对Skills for Real Engineers感兴趣，我也可以考虑单独写一篇文章交流。
BTW，这篇文章我没有使用任何AI工具生成，全部手敲。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Foreword
This article is just a record of my consciousness in the last six months and years using the tools of AI to do my daily work.
In these times, it is very difficult to continue working without the use of AI tools.
Claude Code Terminal
In fact, the first AI tool I used was Trae, because of the company's cooperation, Naho Token gave too little to be qualified for routine tasks, and the Claude Code terminal was used at the beginning of this year, using the model GLM 4.7 (free access model for colleagues) and the MiniMax package bought by his own paid.
What has been done here, actually, in the last article I've already had a lot to say about it, but it's not too much to describe, and you can see it for yourselves. Let's talk about what I've been doing lately. People
Consciousness is that AI can really be more efficient, because it is difficult to provide an overview because all the operational projects are code upgrades, library upgrades, and technological upgrades, and I understood that AI is still not good at particularly complex work.
Later, for various reasons, as well as Claude Code's own death, it became imperative to refrain from using Claude Code.
Codex App Desktopend
And then I used Codex App, and I was very opposed at first. I was used to the Claude Code terminal, and now, in exchange, is it not a cost for me to recreate the usage?
So a couple of teaching videos were painted in Bilibili, and it was operationalized, and it was discovered that there was a Codex App desktop end and what was needed as a bicycle.
Codex App's desktop is good!
With Codex, anything like uploading pictures, files, folders is much more friendly than the terminal.
Apart from programming, using Codex, it can read the whole project more systematically. The user-friendliness of the page UI makes you use it.
And I've been exposed to more models than GLM, Mini Max, and Deepseek is about to learn at this time.
Deepseek made me feel like I was in prison, but there was Code Plan in my hand, and I couldn't use it. And I don't know how the gap between models and models can be so large.
The real turn is in AI's workflow use.
AZworkflow
The previous coding exercise was basically a minor patching of existing codes, or small trajectories, with the use of AI input and output, and the true chapter of the project starting from 0, which is essential if AI is to start from 0.
GSD
The boss of the company was recommending the use of the GSD process internally, and I tried to use it in my work, and it was interesting to say that the GSD was interesting, because after the original author ran the bucket for some reason, he had to set up another stove.
Differences between models found using GSD
I think it's quite a good process in the use of GSD, because Skill is too much, and sometimes I don't know how to choose, and my Code Plan is MiniMax, and in the process, I'm finding that it's really stupid, and it's not going to compress context itself, and it's obviously going to trigger it, and it's going to do it on its own, without Skill, and it's going to do it.
There was no way to use the GLM of a colleague, who had tried top up on Deepseek, and found that prompt, which he had originally written, was a problem simply caused by a model problem.
As a result, the Code Plan of Mini Max drifted, and the replacement of models was necessary for subsequent development.
Skills for Real Engineers
As the project continues and the model evolves, we find that the use of GSD is becoming more and more problematic than it is for us. So after several water tests, our AZworkflow moved to the top of AZworkflow at the front end of Matt Pocock.
Use Skills for Real Engineering
First of all, you need to use this Skill, at least guarantee model capabilities. This stage I myself, top up, took the Code Plan of Codex Plus, in addition to Deepseek, and it's a mess.
And with the mainstream process of Grill me/Grill with Docs to Spec to Pickets to Implement, the first step was Grill.
It's a constant demand input, Grill's 30 questions are a breakdown, and it's a good thing that Skill's upgraded and has a recommended answer, but, but, as a developer, you have the final decision, and the recommended answer may not be what you want, so it must be carefully examined!
The process behind it is simple: ToSpec → → Z Q mp mp Q Q.
I think it's going to be a lot easier to come down, except for some of the Fettokens.
AI, I, and awareness.
It's just a record of my experience, and these are my own thoughts.
I'm still tired at work.
In fact, simply writing codes is not exhausting, because difficult codes are barely written, and technical bottlenecks are virtually non-existent, but as AI becomes more and more capable of writing codes, the boundaries of my own duties have changed, which make it more difficult.
For example, I understand the need to enter into me, and then write the code, but now it's becoming, and I understand how to enter into AI, so that the expected functional code that AI wrote, and the prompt writing out of the brain becomes a problem, and if it isn't an AI process, topec → toTickets, I think it's a waste of brains to write a standard input.
Moreover, it is not easy for me to have to review a large number of documents without writing codes, even at the beginning of the project, and even under a functional module, I would have to check whether the codes I wrote were in line with expectations, although there were various md files and rules to bind the codes, but I had to go through the process.
Moreover, from a leadership point of view, I must have an AI and the workload and schedule are much faster than before. And the company's support for AI is not very friendly, it's actually paid, and since paid is working, I expect to finish my work as planned and leave work on time, but that's not true.
There's an AI. I've changed my mind a lot.
Before, I always thought that it would be difficult for AI, even if it developed fast, to get rid of the mobile end, because it was different from the front end, because it relied heavily on multi-modular models (UI needs), while Xcode's configuration was very complex, but with the development of AI, and even with Xcode27's configuration changed to json, it actually served AI.
From the very beginning, I refused to accept it, and I embraced it positively, and it changed very quickly.
The only anxiety at the moment is what can I do? Is there anything I can do about my position? Will I be killed?
What have you been doing for months?
The former RxStudy was just moving my own project with the company's access.
In July-September, I completed the landing of React Native on both sides of a project; in another project, KMP and CMP on both sides, with the prospect of HarmonyOS Next end, and we even completed a partial validation of Quickly+KMP on HarmonyOS Next end, as well as the full development of Flutter.
As an iOS development, I don't know, I don't know, I don't know, I don't know, but I ended up on the ground, and I don't care about the configuration of the project, I just have to see if it's written or not.
Now everyone calls AI a genius programmer, and at this point, apart from looking at the code structure and hierarchy that it writes, it's hard for me to read it for most of the time, because it's more advanced than I am, and most of the time it's better structured and thought.
With AI, writing codes is no fun anymore.
It is absolutely true that when humans stop exploring, there is no fun, and this is just like playing games, and it's fun to explore one step at a time in the face of an unknown vast map. There's no technical difficulty with the AI writing code, and it seems to me that, as long as Token is in place, the time cycle allows, reasonable needs can be developed.
And we, like the staff above the water line, go step by step to check and check.
Fun becomes how to have a good idea, even an interesting interactive or animated, and when I get down from the heavy and repetitive CV, I get more involved in the input of demand, whether the demand is reasonable and whether it can run out.
AI, on the one hand, modeling is the key.
I've seen a lot of people say there's a difference in the performance of the same model under different ICE, but from what I've used, GLM, Mini Max, Deepseek and ChatGPT, if the model itself is strong enough to fly the bricks.
I'll always use Skill, MCP and plugins, even to learn how to install and use them, but in the course of our development we find that for a model that is strong enough to say what to do, let permission go, and leave the rest to itself.
It is also important to note here that permission has been granted a double-edged sword, which is convenient and also dangerous.
All right, that's some of my thoughts, and you're welcome to talk, and if you're interested in Skills for Real Engineers, I could consider writing a separate article.
BTW, this article I didn't generate with any al tools, all banging.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.