参赛赛道：赛道一 · TRAE 全链路实战 话题：#全新 TRAE 实战体验
用全新 TRAE 一个人做完了整个小工具。最大感受： 它的进步不是"更会写代码"，而是"能把一件事推到底" ——Agent 模式丢整段需求直接回一个能跑的应用，我负责验收和精修；视觉和演示态这种我心里有数的活儿切回 IDE 手动改。以前我是带着 AI 写代码，现在更像带一个能干的实习生做项目：我管判断，它管落地。 启动界面如下：
登录账号授权：
一、先说清楚我要解决什么事
每周五下午四点，我们公司会准时上演同一幕：一群人盯着周报系统的空白文本框发呆。
我这周干了什么？——修了 3 个线上 bug、把首页加载从 4 秒压到 1.8 秒、写了两份技术文档、带新人过了代码、开了 5 个会。
写进周报变成这样：
1. 修复线上bug
2. 优化首页加载
3. 编写技术文档
4. 带新人
5. 参加需求评审会
领导看完的反应大概是："这人这周挺闲的。"
但真相是我忙到没空喝水。
问题不在工作量，在表达能力。 同样一件事，写成"修复线上bug"和写成"主导核心系统稳定性治理，累计排查3项线上关键缺陷，故障率下降，保障业务连续性"——后者才是升职加薪那个人写的东西。
这个转化过程有个专业名字叫"述职语言包装"，通俗点说叫"把执行语言翻译成管理语言"。它不难，但极其消耗心力：一句一句想措辞，平均一条要磨 5 到 10 分钟，一份周报写下来 40 分钟起。
于是我决定：这件事必须自动化。而且不能靠大模型每次联网现编，得有一个稳定的、可离线运行的翻译引擎。
这就是我用全新 TRAE 从 0 到 1 做出来的东西——「打工人战绩翻译机」。
对话如下：
二、关键决策：我的第一版方案被自己推翻了
这段我想单独讲，因为它是这次开发里最有价值的思考，也最费时间。
第一版我做的不是这个工具。 我当时想蹭 2026 年的热梗"夹角打工人"（形容卡在内卷和躺平中间的人），做了个「夹角打工人诊断仪」——问卷答题、向量余弦夹角算分、出诊断报告。
技术上挺酷：真的用了 cos θ = (a·b)/(|a||b|) 算夹角，还配了雷达图和极坐标向量图。
但我做完自己用了一遍，发现它是个娱乐产品，不是一个解决问题的工具。
- 诊断完，用户情绪被调动了，但明天上班该写周报还是不会写
- 它解决的痛点强度不够：好玩，但不高频、不刚需、不痛
- 参加"真实工作场景提效"赛道，它偏题了
我推翻它，重新选了"周报/述职"这个方向，理由是三个字：真、频、痛。
这是本次开发第一个关键决策，也是我踩的第一个坑：好玩 ≠ 有用。 参赛要的不是酷，是"用完了问题真的被解决"。
三、全链路开发过程：Agent 模式推进 + IDE 模式接管
这一段是我的真实操作记录。全新 TRAE 的 Agent / IDE 双模式，在这次开发里的分工特别明确，我基本是这么用的：
环节1：让 Agent 帮我造"翻译规则库"（发散型任务）
工具的心脏不是 UI，是规则引擎——什么关键词该翻译成什么话。这一步最费脑，因为要覆盖开发、产品、设计、运营四类岗位的表达习惯。
我在 Agent 模式里给的第一条指令是这样的：
我要做一个"打工人周报翻译机"工具，核心是把大白话工作内容翻译成高大上的述职语言。请帮我设计一套翻译规则库，要求：
- 覆盖开发、产品、设计、运营四类常见岗位；
- 每条规则包含：触发关键词（多个）、翻译模板（含数字占位符）、分类标签、分值权重；
- 覆盖这些场景：修bug、写代码、优化性能、写文档、开会、做需求、带新人、测试、上线、处理用户反馈、数据分析、设计页面、运营活动、摸鱼；
- 翻译模板要"有梗但不离谱"——能让人会心一笑，但确实也说得通；
- 输出成可直接转成 JavaScript 代码的结构化格式。
Agent 一口气生成了 15 条规则。
这里有个关键调优，值得单说： 我第一版给的指令里没有第 4 条"有梗但不离谱"。结果它生成的模板全是"推进XX工作，提升XX效率"这种公文套话——正确，但不好玩，用户不会想截图分享。
加上调性约束之后，出来的东西质感完全变了。比如"摸鱼"这条，它给的原始版本是"进行精力恢复"，我觉得太正经，手动参与了润色，最终定稿：
{
keywords:['摸鱼','发呆','刷手机','带薪','上厕所'],
category:'战略级摸鱼', color:'#ffd23f', score:2,
template:(n,raw)=>
`进行<span class="num">深度战略级</span>精力恢复与创造性思维孵化，`+
`为后续高强度产出蓄能。<span class="hl">（摸鱼也要写出格局）</span>`
}
经验：给 AI 下指令，"调性约束"比"格式约束"更能决定产出质量。 格式约束决定它有没有跑对，调性约束决定它跑出来好不好用。
环节2：一次生成完整单文件应用（Agent 一把推下去）
规则库拿到手后，我把规则 + 交互需求一次性给 Agent，让它整份实现。
指令大意：
基于以下翻译规则库，做一个单文件 HTML 交互应用：
- 用户输入本周工作内容（一行一条，大白话）；
- 提供 4 个岗位示例（开发/产品/设计/运营），点击即填充；
- 点"翻译"后输出四块：战绩评分卡片、逐条翻译结果（原话→翻译，带分类标签和分值）、自动组装的完整周报段落（含本周总结/关键数据/下周计划）、职场黑话等级评定；
- 支持一键复制周报全文；
- 暗色主题、移动端适配；
- 技术约束：纯 HTML+CSS+JS 单文件，不依赖外部库，离线可跑。
它直接产出了一个可运行的单文件应用。这一步是我觉得全新 TRAE 变化最明显的地方——不是给我一堆代码片段让我拼，而是直接给出一个能打开就跑的产物。我拿到之后只需要验收和精修。
环节3：切回 IDE 模式做视觉精修（这里必须手动）
Agent 给的版本功能对了，但视觉平庸。我切到 IDE 模式手动改了三处：
- 数字高亮：翻译结果里的数量词用青色、关键结论用金色，让"述职感"从视觉上先立住；
- 评分上限设为 150：防止条目太多分数爆炸，同时保住"传奇战绩"这个顶级的稀缺感；
- 暗色主题统一：把 Agent 用的几套杂色收敛成 5 个语义色（粉=功能、青=数据、金=结论、绿=成功、红=风险）。
双模式的真实体会： 发散、重复、需要一次性推到底的活儿交给 Agent；需要精准控制、我有明确审美的活儿必须切回 IDE。切换成本几乎为零，这一点比以前用其他工具来回倒腾要顺。
环节4：把它变成一个"打开就能截图"的演示态（踩坑修复）
应用做好后我遇到一个很实际的问题：打开是空的。 得先点示例、再点翻译，才能看到成果——对演示和截图很不友好，别人第一眼看到的是空输入框，不知道这工具能干什么。
改法很简单，三行代码——让页面加载后自动填示例并自动跑一遍：
window.addEventListener('DOMContentLoaded', () => {
document.getElementById('inputBox').value = EXAMPLES.dev; // 自动填开发岗示例
updateCharCount();
setTimeout(() => translate(), 300); // 自动翻译，直接出成果
});
这个坑很小，但我想强调它：做工具和做作品是两件事。工具追求"功能完整"，作品追求"第一眼就懂"。 自动演示态属于后者。这一步我放在 IDE 模式里改，因为需求太明确了，描述给 Agent 反而更慢。
四、落地成果：一个能真跑、能截图、能复用的工具
4.1 成品是什么
一个单文件 HTML 应用，双击打开即用：不需要后端、不需要装依赖、不联网也能跑。因为它是纯规则引擎，不调用任何大模型 API，响应也是瞬时的。
4.2 功能清单
4.3 实际运行结果
运行结果示例（我用开发岗示例实测的翻译前后对比）：
4.4 效率对比（量化）
核心结论：单次写周报从 40 分钟压到 2 分钟，效率提升约 95%。 一年按 48 周算，一个人一年能省下约 30 小时——而且省下的是最消耗心力的那部分。
4.5 顺带复用了 TRAE 的两个平台能力
除了双模式，我还用了两个新能力，都是顺手用上的，不是硬凑：
- 我的产物统一管理：这次开发产出很杂——demo 的 HTML、规则库代码、几版设计稿截图、还要写的文章。它们集中放在一处，回头找"那版带摸鱼彩蛋的规则库"时不用翻聊天记录，直接就能定位到。
- 模板库起手：这个工具是纯前端单文件，结构上有固定套路（输入区 → 触发按钮 → 结果分区 → 复制导出）。我没有从零搭骨架，而是拿了一个相近的单页应用模板起手，把主要精力放在规则引擎和内容上——真正有价值的部分是规则库和调性，不是脚手架。
五、踩坑清单（这四条比功能列表更有用）
坑1：第一版方案选错，白干一轮。 我一开始做的是"夹角打工人诊断仪"，好玩但无用。教训：参赛/提效类项目，先问三个问题——它真实存在吗？高频吗？痛吗？三个都是"是"才动手。
坑2：指令没给调性约束，产出全是公文套话。 "帮我设计翻译规则"和"帮我设计翻译规则，要有梗但不离谱，让人会心一笑"，产出差一个量级。前者给"推进工作提升效率"，后者给"深度战略级精力恢复与创造性思维孵化"。
坑3：没做兜底逻辑，输入命中不了规则就输出空。 15 条规则不可能覆盖所有输入。我加了一条默认规则：没匹配到就输出"推进 N 项日常工作任务的执行与落地，保障业务运转与团队协作顺畅"。兜底不是可选项，是必选项。
坑4：多重关键词同时命中，会误判。 输入"开会影响优化方案的会"同时命中"开会"和"优化"。修复方式是匹配前按关键词长度排序，让更具体的关键词优先命中——这个词序修复我是把现象描述给 Agent，让它自己改的，比我自己 debug 快。
六、可复用的方法论（三张模板卡，换词就能用）
模板卡1：规则引擎设计（适用于任何"输入→翻译/转换"类工具）
我要做一个{工具名称}，核心是把{输入类型}翻译成{输出类型}。
请帮我设计翻译规则库：
1. 覆盖 {N} 个常见场景：{场景列表}
2. 每条规则包含：触发关键词（多个）、翻译模板（含{变量}占位符）、分类标签、分值权重
3. 翻译调性：{有梗但不离谱 / 严肃专业 / 活泼幽默}
4. 输出成可直接转成代码的结构化格式
照搬案例：简历优化器（大白话经历→专业简历语言）、邮件润色器（口语→商务正式）、合同翻译器（日常表述→法律条款语言）。
模板卡2：关键词匹配 + 数字提取代码模式
// 关键词优先匹配：关键词更长的规则优先，避免误触发
function matchRule(text){
const rules = [...RULES].sort((a,b)=>
Math.max(...a.keywords.map(k=>k.length)) -
Math.max(...b.keywords.map(k=>k.length))
);
for(const rule of rules)
for(const kw of rule.keywords)
if(text.toLowerCase().includes(kw.toLowerCase())) return rule;
return null; // 无匹配 → 走兜底
}
// 数字提取：阿拉伯数字 + 中文数字都支持
function extractNumber(text){
const m = text.match(/(\d+)/);
if(m) return parseInt(m[1]);
const cn = {'一':1,'二':2,'两':2,'三':3,'四':4,'五':5,'六':6,'七':7,'八':8,'九':9,'十':10};
for(const [k,v] of Object.entries(cn)) if(text.includes(k)) return v;
return null; // null 时模板用默认值
}
模板卡3：交互式工具的"仪式感"流程
1. 输入区：提供示例标签 → 降低使用门槛
2. 一键触发：一个按钮完成全部处理
3. 评分卡片：大数字 + 等级 → 制造冲击力
4. 逐条结果：原话与翻译并排 → 建立可信度
5. 组装输出：把零散结果拼成可直接交付的文档
6. 一键复制：拆掉"用起来"的最后一道门槛
7. 自嘲评级：留一个让人想截图分享的彩蛋
用全新 TRAE 把这周的周报工具从 0 推到能用，说几个最真实的感受。
最明显的变化不是"AI 更聪明了"，而是"它开始能交付了"。 以前用 AI 编程，拿到的是一堆片段，我还得自己拼；这次我在 Agent 模式里丢了一整段需求（规则库 + 交互流程 + 技术约束），它直接给回来一个打开就能跑的单文件应用，我的工作从"写代码"变成了"验收 + 精修"。这个心理落差挺大的——省下的不是敲键盘的时间，是那种"一件事推不到底"的疲惫感。
第二个感受是双模式的分工真的成立了。 我的实际用法很朴素：发散、重复、需要一把推下去的活儿丢给 Agent（设计翻译规则库、整份前端实现、定位边界 bug）；需要精准控制、我心里有明确答案的活儿切回 IDE 手动改（配色语义、评分上限、"打开即演示"那三行）。切换成本几乎为零，这点比以前在工具之间倒腾舒服太多。
第三个是意外收获：产物统一管理救了我。 这次产出特别杂——HTML、规则库代码、几版截图、还要写的稿子。集中放在一处之后，我不再靠翻聊天记录找"那版带摸鱼彩蛋的规则库"，直接定位。这种"东西都在"的踏实感，比任何单个功能都实用。
总结
周报这件事，技术含量不高，但极其消耗心力。每周一次，全公司加起来是个巨大的时间黑洞。
我做这个工具不是教人吹牛。它服务的是那些 "活儿干得漂亮但说不出口"的人——你的工作值得被看见，你只是缺一个翻译。
一个人干活，最怕的不是做不出来，是想不清楚。 这次的体会是：Agent 模式帮我把"想"的部分快速铺开成可验收的东西，IDE 模式让我在关键处保留精确控制权，两者来回切，一个人也能把一条链路跑完。
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。
# New TRAE Experience
With a brand-new TRAE one person finished the whole little tool. It's not "better code," it's "getting one thing to the end" -- the Agent model throws the whole demand straight back to a running app, and I'm in charge of acceptance check and precision; the visual and demonstration kind of work that I have in mind is going back to IDE manually. I used to write with an AI code, and now it's more like taking a capable intern for the project: I judge, it falls. The interface is as follows:
Login number authorizes:
First, let's get this straight.
Every Friday at 4:00 p.m., our company will be doing the same thing on time: a group of people staring at blank text boxes in the weekly newspaper system.
What have I done this week? -- 3 line bugs have been fixed, the front page has been loaded from 4 to 1.8 seconds, two technical documents have been written, new people have passed the code, five meetings have been opened.
This is what happens in the weekly newspaper:
1. Uplink bug
2. Optimizing loading of the home page
3. Preparation of technical documentation
4. Bringing new recruits
5. Participation in needs assessment sessions
The leader's reaction after reading was probably, "This guy's been busy this week."
But the truth is I'm too busy to drink.
The problem is not the workload, but the ability to express. The same thing, written as "bug on the restoration line" and "dominant core system stability governance, cumulatively sorting out three critical fault lines, decreasing failure rates and ensuring business continuity" — the latter was written by the person who had been promoted.
This conversion process has a professional name called "Speech Language Packing" and it's called "Translating Implementation Language into Management Language." It's not hard, but it's very hard: it's got to be five to 10 minutes on average, 40 minutes in a weekly.
So I decided that it had to be automated. Moreover, LLM cannot be built on each connection, but has to have a stable translation engine that can be run by offline.
This is what I've made with a brand-new TRAE from 0 to 1 -- the Worker Performance Translator.
The dialogue was as follows:
Key decision-making: My first version was overturned by myself
I would like to say this paragraph alone because it is the most valuable thought in this development and the most time-consuming.
This isn't the first version I do. I was trying to tumble the 2026 hot-tip-tip-tip workers, making a "tip-tip-tip-tip-tip-tip-tip-tip-tip-tip-tip-tip."
Technically cool: It really uses cos = (a.b)/(a||b|) to calculate the accelerator with radar maps and polar coordinate vector maps.
But I did it myself and found it to be an entertainment product, not a problem-solving tool.
- After the diagnosis, the user's emotions were moved, but tomorrow's work should be done in the weekly. Write
- It doesn't solve pains enough: fun, but not high frequency, not demand, not pain.
- It's off the subject.
I overturned it and re-elected the "weekly newspaper/statement" on the grounds of three words: truth, frequency, pain.
This is the first key decision of this development, and the first pit I've stepped on: fun Zip is useful. It's not about cool, it's about "no problem really solved".
III. Full chain development process: Agent mode advance + IDE mode takeover
This is my real operating record. The new TRAE's Agent / IDE double mode, the division of labour in this development is particularly clear, and I basically use it:
Step 1: Get Agent to help me build the Translation Library.
The heart of the tool is not UI, it's the engine of the rule -- what the keyword is to be translated into. This is the most costly step, since it covers the expression habits of developing, producing, designing and operating four types of jobs.
Here's the first instruction I gave in Agent mode:
I'm going to make a tool called the "Worker's Weekly Translator" that will focus on translating the content of the work into the high-profile narrative language. Please design a translation library for me.
- To cover the development, product, design and operation of four common categories of jobs;
- Each rule contains: trigger keywords (multiple), translation templates (including digital placeholders), classification labels, weighting;
- Coverage of these scenes: bug repair, code writing, performance optimization, document writing, meetings, needs, new people, testing, online, user feedback processing, data analysis, page design, operational activities, fish mapping;
- The translation template must be "unbelievable" - it makes people laugh, but it makes sense;
- Output into a structured format that can be directly converted to JavaScript code.
Agent created 15 rules in one breath.
Here's a key modulation, worth saying: there's no fourth rule in my first edition. As a result, all the templates that it produces are the post kit called "Advance XX Work, Enhance XX Efficiency" -- correct, but not funny, and users don't want to copy and share.
And with the modulative constraints, the quality of what comes out is completely different. For example, the fish toucher, which gave the original version of "Rehabilitation," I thought it was so serious that I was manually involved in the decorating that it was finally finalized:
{
Keywords: ["Sniff", "Swipe the phone", "pay", "to the toilet"],
Category: 'Strategic fishing', color:'#ffd23f', score:2,
template:(n,raw)=>
<span class=num" > </span's energy recovery and creative thinking incubation, >
`Accumulation for subsequent high-intensity output. <span class >
}
Experience: Directive AI, "modular constraints" are more likely to determine output quality than "format constraints". Format constraints determine whether it's running right, and modem constraints determine whether it works.
Step 2: Generate a full single file application at a time (Agent pushes)
Once the rules library is in hand, I give the rules plus the interactive demand to Agent once and for all.
Instructions to:
An HTML interactive application based on the following translation rules library:
- User input for this week ' s work (one line, big words);
- Provision of 4 examples of jobs (development/product/design/operation), click on fill;
- Four pieces of output after "translation": performance score card, article-by-article translation (translation, with classification labels and fractional values), full weekly paragraphs of automatic assembly (including summary of the week/key data/plan for next week), job-level black language rating;
- Support for a one-click copy of the full weekly paper;
- Dark themes, mobile end appliance;
- Technical constraints: Pure HTML+CSS+JS single file, not relying on an outer library, offline can run.
It directly produces an operational single file application. And this is the one thing that I think is the most significant change in the whole new TRAE -- not to give me a bunch of code pieces that I can spell, but to give a product that can open and run. All I need after I get is acceptance check and precision.
Step 3: Cut back the IDE mode for visual refinement.
Agent gave the right version, but the visuals were mediocre. I cut the IDE mode manually by three:
- High-profile numbers: the number of words in the translation results is green, the key conclusions are gold, so that the "confessional sense of duty" is seen first;
- The ceiling for rating is set at 150: prevent the entry from going off with too many points, while preserving the sense of scarcity at the top of the title, "Traditional performance";
- Dark theme harmonization: Consistency of Agent's collections into 5 semantic colours (powder = function, green = data, gold = conclusion, green = success, red = risk).
The real experience of the two models: dispersing, repeating, one-time push-through work to Agent; precise control to control, and Idea to get back to. The cost of switching is almost zero, which is better than the previous use of other tools to reverse.
Threshold 4: Turn it into a demonstration of "open and clear."
When the application was done, I had a real problem: it was empty. It takes a few examples and a few translations to see the results — it's very unfriendly to presentations and screenshots, and the first person to see an empty input box is not sure what this tool can do.
The change is simple, three-line code -- to automatically fill out examples and run them over once the page is loaded:
window.addEventListener('DOMContentLoaded', () => {
Other Organiser Example
updateCharCount();
setTimeout(() = > translation(), 300); / / Autotranslation, direct result
});
The pit is small, but I want to emphasize it: it's two things to do with tools and works. Tools for functional integrity, works for first sight. Auto-demonstration belongs to the latter. I changed this step in the IDE model because the demand was so clear that the description to Agent was slower.
IV. Landfall outcomes: a tool that can run, capture and reuse
4.1 What's the finished product?
A single file HTML application, double-click to open it: does not need a backend, does not need to be attached, cannot run without a connection. Because it's a pure rule engine, without calling any LLM API, the response is instantaneous.
4.2 List of functions
4.3 Operational results
Example of run results (I use a translation comparison based on the development of a sample):
4.4 Efficiency comparison (quantization)
Core conclusion: Single weekly newspaper presses from 40 minutes to 2 minutes, increasing efficiency by about 95 per cent. In 48 weeks a year, a person can save about 30 hours a year - and save the most exhausting part.
4.5 Capacity of two platforms where TRAE was used
In addition to the two models, I have two new capabilities, which are hand-held, not hard:
- Integrated management of my products: this development is a multiplicity of outputs — demo HTML, code of rules, several editions of scripts, articles. They're concentrated in one place, and they can be located directly without turning over the chatting records when they look back at the "The Rules Library with FishEggs."
- Template library starter: This tool is a purely front-end list file with a fixed set of paths in its structure (Input Area Trigger buttons, result partitions, copy export). I didn't start with a single sheet, but I started with a similar template, focusing on the rules engine and the content. - The real value is in the library and the tone, not scaffolds.
V. List of pedestals (four more useful than functional lists)
Pipe 1: The first version was wrong. I started with the "Playing Workers' Diagnostics" and it was fun, but it didn't work. Lessons learned: For the competition/effect-type project, three questions first — is it real? Is it high? Does it hurt? All three are "yes."
Pipe 2: The instructions are not transposed and the output is a clerical package. Help me design the rules of translation, help me design the rules of translation, make people laugh, and output is down a scale. The former gives "promoting work to improve efficiency" and the latter gives "incubation of deep strategic energy and creative thinking".
Pit 3: Without a bottom logic, enter the rules and output empty. 15 It is not possible for the rule to cover all inputs. I added a default rule: "Putting N tasks forward and landing without matching, ensuring smooth operations and teamwork." The bottom is not an option, but an option.
Pipe 4: Multiple keywords hit simultaneously and are miscalculated. Enter the words "meeting to influence the optimisation program" and "optimization". The fix is sorted by the length of the keyword before matching, giving priority to more specific keywords -- the word that fixes me by describing the phenomenon to Agent and making it change itself, faster than myself.
VI. METHODOLOGY FOR RESTRUCTION (three template cards, reworded)
Template card 1: Rules engine design (for any input translation/conversion tool)
I'm going to make a {tool name}, and the core is to translate {input type} into {output type}.
Please help me design the translation library:
1. Overwrite {N} Common scene: {Scene List}
2. Each rule contains: trigger keywords (multiple), translation template (including {variant} placeholders), classification labels, split of values Heavy
3. Translator tone: {invasive but not arbitrary / serious professional / lively humor}
4. Output into a structured format that can be directly translated into codes
Replicating cases: CV optimizer (presumable language of biographical experience), mail monetifier (official commercial language), contractual translator (normal language of legal provisions).
Template card 2: Keyword Match + Digital Extract Code Mode
// Keyword priority matching: longer keyword rule priority to avoid error triggers
function matchRule(text){
const rules = [...RULES].sort((a,b)=>
Math.max(...a.keywords.map(k=>k.length)) -
Math.max(...b.keywords.map(k=>k.length))
);
for(const rule of rules)
for(const kw of rule.keywords)
if(text.toLowerCase().includes(kw.toLowerCase())) return rule;
/ No match
}
/ / numeral extraction: Arabic + Chinese numerals support
function extractNumber(text){
const m = text.match(/(\d+)/);
if(m) return parseInt(m[1]);
Const cn = {1,2,2,3,3,4,4,5,5,6,7,7,8,9,9,10;
for(const [k,v] of Object.entries(cn)) if(text.includes(k)) return v;
The template is defaulted when reverting Null; / / Null
}
Template card 3: Ceremonial sense process for interactive tools
1. Entry area: provide illustrative labels lower use threshold
2. One-click trigger: a button completes all processing
3. Rating card: big + grade manufacturing shock
4. Article-by-article results: original language in parallel with translation
5. Combining output: collating scattered results into directly deliverable documents
6. Copy one key: remove the last threshold of "used"
7. Self-scribbling rating: leaving a coloured egg that makes you want to share
Use a brand-new TRAE to push this week's weekly paper tool from zero to use, say a few of the most real feelings.
The most obvious change is not "AI is smarter," but "it's starting to deliver." I used to program with AI, I got a bunch of pieces, and I had to spell it myself; this time I lost a whole section of the Agent model (rule library + interactive process + technical constraints), and it just gave back a single file application that opened up and ran, and my job went from "writing code" to "acceptance check + fine." It's kind of a mental difference -- it's not the time to knock on the keyboard, it's the sense of fatigue of "one thing goes nowhere."
The second feeling is that the dual division of labour has really been established. My actual usage is simple: dispersing, repeating, requiring a push-over job to go to Agent (design translation rulebook, whole front-end realization, positioning boundaries bugs); and cutting back to IDE (colored words, rating ceilings, "open and show" lines) that require precision control and clear answers in my mind. The cost of switching is almost zero, which is far more comfortable than it used to be between tools.
The third is the windfall harvest: Unified management saved me. This output is particularly complex — HTML, code of the rules library, several editions of the cut-off chart, and additional manuscripts. After concentrating in one place, I stopped looking for "the rules house with fish poaching eggs" by flipping the chat. It's a feeling that everything is real, more practical than any single function.
Summary
The weekly report on this is low-tech, but extremely hard-working. Once a week, the whole company adds up to a huge black hole of time.
I'm not teaching people to brag. It serves people who do "good work but can't say it" -- your job deserves to be seen, you just lack an interpreter.
It's not that you can't do it, it's that you can't think. And this time, the Agent model helps me put the part of the thought fast into something that can be acceptance check, the ICE model allows me to keep precise control of the key, and the two cut back and forth, and one can run a link.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.