EP312 – AI in 2026 Worker's research on metallurgical engineering and economics Buying, selling, selling, selling, selling, buying and selling!

There are a bunch of new AI tools popping up every day. Do you often don’t know where to start? Don’t worry, here is the AI Lazy Newspaper to help you select the five most discussed AI tool videos every day. For ten minutes a day, let AI enter your life, and you will become an efficiency master without knowing it. I think what I am going to talk about today is really to my taste. Have you ever thought that AI doesn’t just help you create a program for you? It actually helps you with animations and uses the keyboard to get things done. And today there is a super practical method that allows you to use AI to make a complete application on the weekend without having to rewrite every line of code. Let's start directly. Let's talk about Manis first. The most impressive thing about it is that it will really help you organize your computer. The biggest difference between Manis and the cloud AI tools you usually use is that it runs directly on your computer. It has two very critical permissions, one is to access your local files and the other is to directly execute commands on your operating system. It is not a brain that lives in the cloud, it is a worker that lives in your computer. Imagine that you are a person who takes a case, or a computer with a bunch of messy ordinary people. You have a folder with 300 publications in different formats. Under normal circumstances, you might ask an AI to help me write a script to rename these files. Then you open the mid-range machine and find that the environment is not right. After working on the entry problem for a long time, you finally finished it, but you are tired of using Manis. If you do, just point the folder to it and rename all the files for me in the format of date plus customer name. It will do it directly. It will scan each file, understand the file content, and then execute the file system instructions on the local machine. It will be done in a few seconds. But what really scares me is its ability to develop applications. You can ask it to help you build a desktop application from scratch. You describe what you want, and it will write the code. It will search the local environment, debug the errors encountered, and then package the application. Everything is done automatically on your computer. In the past, it may have taken a development team several days to build something. It can be done in a few minutes. My personal favorite feature is remote control. Manis allows you to control your computer with your mobile phone. You leave your laptop open at home in a cafe and suddenly think of a file you want to use. You open the application on your phone and ask Manis to help me find the report file for the third quarter on the desktop and send it to me. Your computer becomes a remotely controlled AI node. I know, it sounds a bit scary to give such a large authority to AI. You really need to pay attention to security issues. I suggest you run it in the test environment first and don't let it encounter important data at the beginning. But just help me make thumbnails of these original video clips and organize them into a new folder, and leave a text file to list the links. It feels really good to run from beginning to end with one command. It will be very painful to go back and click manually. Speaking of voice AI, the next tool will solve another very troublesome area. If you have recently been engaged in the development of voice AI. You know that Tone. It looks very simple on the surface. Convert speech to text. Throw it to AI and then convert the response into speech. That's it. But a real phone call is a mess. The user will interrupt you, will be suddenly silent, and will ask when the AI is talking about the budget. Are you a robot? In order to deal with these situations, most people use off-the-shelf platforms like VIP, Blend Retail, etc. They are fast to use, but you will cry when the bill comes. You are paying for the usage of large language models, speech generation fees, telecommunications fees, and the platform itself. And the most troublesome thing is that you don’t know where the problem is. Dobra is an open source voice AI platform designed for developers who want to set up and control it by themselves. It supports Docker deployment. You can grab the code and run a command to get in. But what makes it smart is not just the open source thing. It also has a construction tool that can drag and drop the process. You can draw the entire conversation process, prompt nodes, qualification review steps, call external services to update the customer database, and branch to live customer service. For engineers, what really shines is its tracking function. You run a test call. For example, a customer calls in and the robot asks about the company's size and budget. Dobra will give you a complete transcript of the conversation, delay tracking, records of each tool call, and actual call recordings. Not only do you know whether the robot ran, you know why it ran like this or when the user suddenly sneezed. This project is still very early, but it is in the right direction. Write code where it really matters, use visual tools to handle the conversation flow, and then build it yourself. When you scale up, you will not be exhausted by the pricing of the platform. Controlling the infrastructure yourself is important for formal products, but if it is just your own personal daily process, what you need is a helper who will always accompany you. This brings up the latest update of Hermes. Hermes' goal is to create a persistent autonomous AI that can help you monitor the system in the background. But he recently solved I think one of the most annoying problems in using AI every day is the problem of memory fragmentation. Think about what your current settings look like. Your Cloud Code lives on your desk and helps you do the heavy work. It knows all your code libraries and your programming habits. Then your mobile assistant may be Hermes running on Telegram. You throw ideas into it while walking the dog. The problem is that these two brains never communicate. You upload an idea to Hermes on your mobile phone, but you still have to start with Core before you return to the desktop. Explain the whole concept. There is no handover at all. This Hermes update adds a layer called Agent Tye COS. Connect the Hermes Telegram robot directly to your Cloud Code Core. It will track everything you do in this episode, monitor your usage charges on different models, record your conversations, and even has a function where it will dream at night, that is, it will analyze your interactions that day and then suggest directions for process improvement. Hermes can now directly access the same memory library. They also introduced a feature called Pantheon's thing is a custom AI personality system in Hermes. Each personality has a designated model and role. For example, you want to do an in-depth research on marketing strategies, but you don't need to burn the expensive fixed model usage fee for this. You can summon a personality, for example, call him Labyrinth. This personality is set to use a cheaper model, such as Deep Seek or a certain Lama version, specifically for in-depth research. You send a message to Hermes on Telegram to hand it over. Labyrinth helped me find five new customer development strategies for roofing companies. Hermes routed the request to the right personality, using the right model, and threw the research results into your Obsidian database. Right. He connected directly to Obsidian, so you were drinking coffee outside. Using Telegram and the Power API obtained by Hermes, he grabbed the contact information of ten local roofing companies. Helped me draft a development letter. Hermes executed the API call, drafted a new one through the integration of CAP, and saved the entire brief into your Obsidian. By the time you return to your laptop, Clarcode already has the complete background of these potential customers and can directly help you create the front-end dashboard of your new application. You can delegate the boring research and data collection work from your mobile phone to a dedicated AI personality, and let the AI on the desktop be responsible for the execution. This is the ultimate lazy character history setting. Speaking of Cloud Code, after listening to Anthropic’s internal engineers share how they use it, I discovered that the way most people use it is fundamentally wrong, including myself. You must have had this moment. You spent thirty minutes carefully typing a super perfect prompt, defining various details. You pressed send and the AI produced something pretty good. But after two prompts, it completely forgot the initial instructions. You use a very fragile prompt to manage a complex process. This simply doesn’t work. Anthropic engineers have four rules, and they have nothing to do with writing longer prompts. The first rule is a prompt skill package, not a prompt. Most of Cloud software development is repetitive work, so Anthropic Engineers build Cloud skill packs. A skill pack is basically a A folder to package the operational knowledge of a certain task. You don't have to re-explain your tone and style every time you want to draft a letter. You just enter a command and throw the original text in. You start to think in terms of the application layer. BundCode creates reusable gadgets. The second rule is that a skill pack is more than just a reminder in a folder. A good skill pack has three levels. The first level is description. This is what Cloud uses to judge whether he should use this skill pack. The description is precise enough. Cloud It can even be automatically triggered without your special request. The second layer is the command, which is a step-by-step operation manual. The third layer is the tool, which is the program script, API call reference file. This is where the real power lies. For example, if you want to check whether a domain name has been registered, you give this skill package a custom shooting script, which can directly query the domain in a programmatic manner. In this way, isn't Cloud guessing that he is executing the program? The third rule is to build small skill packs that can be combined. Don’t build a big monster. My instinct is to build a very large skill pack. Get the idea in one go. Scripting and community copywriting. The team at Anthropic said this is a trap. You want to change the format of Lintint posts. You go to edit that huge skill pack and suddenly your YouTube script breaks. What they do is build small focused skill packs because the Cloud can automatically coordinate between them. Which skill pack is broken, you immediately know where the problem is. There is also a very smart optimization method. Store the script in the skill package. The code is deterministic AI. If Cloud writes a Python script to format the slideshow, don't ask Cloud to rewrite it every time. Save the script in the tool package of the skill package. Next time, Cloud will directly execute the script. You use cheap deterministic native computing to replace the expensive and uncertain AI usage fee. The fourth rule: Tips should become smarter with each use. If you type a normal prompt in the chat interface, the window will always disappear. If you use the skill pack The skill pack will always remain with Anthropic engineers. Treat the skill pack as an asset that will grow unprofitable. If Cloud makes a mistake, they will ask Cloud. Review the back and forth we just had and update your own skill pack instructions so that you don’t make the same mistake again in the future. After using Cloud for 30 days, it should be much better than the first day, because every edge case and personal preference has been permanently written into the architectural strength of the skill pack. This kind of thinking of building modular and predictable systems brings out the last topic of today. How to really use AI to build applications. The term 5coding became popular at the beginning of this year, which is the way you describe what you want and let the AI do it without looking at the code. If you have tried this on a project that is more than a toy level, you know what the results are. The first three functions feel like magic. Speed without direction is a disaster. The best balance point is the developer community. The core of practical vibe coding now is your project and a file in the directory. Agentsmd, tools like cursor, windsword, and clockcode, read your project files to get the background. If you don't have a central rule book, you have to retype your technical station and folder rules in every prompt. One time you forgot to mention that you used sustain for state management, so the AI installed redux by itself. Then your code base suddenly has a split personality agentsmd solves this problem. He starts with one role. You are a senior react native engineer. You put clarity before abstraction. This sentence becomes the filter for generating code for each line. Then you define the technical station. Export native1 does the style. Cursor does the authentication. Sustent does the state management. One line per line. This kills the AI. If the AI keeps making the same mistake, you add a rule in agentsmd. Exceptions. Safe area view uses the standard style. Use native1 class for other locks. You solve it once and record it so that the ai will not make the same mistake again. After agentsmd is locked, you enter the prompt stage. The prompt structure of pract go by coding has four parts. Each prompt has four parts. The first part is to read the agentsmd first and strictly abide by this force context window. Re-enter the rules. The second part is a single task. It is not a good novice guide. Set up authentication and record it to the homepage. It is just the interface of the novice guide screen. The third part is behavioral restrictions. You tell the AI that it is not allowed to touch those things. Do not change the existing verification process. Do not use it in the application. Exposure of Confidential Consulting You are protecting code that you can already run. Part 4 Design Reference. Paste a screenshot. AI immediately understands the visual layout and saves you the time of explaining three-segment spacing and spacing. Use this framework. Take identity verification as an example. You don’t want the AI to invent a custom security mechanism by itself. You want to explore Curk. You save the latest CurkExpo file as a markdown file. Tips close to you say Use real Curk authentication to replace the simulated verification process. 10 Electronics has 9OTP and Google After Auth succeeds, you will be directed to the homepage to use the provided file. Because AI has read your Agent3D, it knows to use your existing Native Y style. Because you have provided the latest file, it will not use the training data to process the expired API. Within a few minutes, you will have a formal level of security authentication. The real test of this process is to integrate complex real-time functions, such as voice and AI tutoring. The goal is to have a real-time low-latency AI language tutor dialogue. You use Stream's React Native SDK for audio transmission and then connect to an AI teacher. Think of it as an AI that listens to you in real time and responds based on the course objectives. You don't want to expose confidential information to the application, so you use the Expo backend path as a secure usage provider. The application asks the server for a pass. The server generates the pass. The application joins the call. You prompt the AI to do the Street Messenger call settings. Use the Expo backend path for pass generation. Don't expose secrets. The interface is loading to add and error status. After the audio transmission is running, Then you start an independent Python service seat. AI teacher, you tell AI to create a Python service and use the OpenAI real-time function. For the AI language teacher, he must read the selected language and course objectives from the call data. Suddenly you open the app on your phone and order a course on Spanish greetings. You are having a real-time voice conversation with an AI tutor who knows which words you should learn. If the AI teacher keeps rushing to talk because of color problems, write a targeted correction prompt. Remove the camera button. Create a medium-sized hold-to-talk microphone button. When I hold the button, I immediately mute the AI and listen to me. The user experience instantly feels like a well-done one. It's really like a production app. Finally you add Posthawk in for analysis, use their AI genie to scan your Reacognitive code base, automatically wrap your app in an analytics provider, and start tracking screen views. You add a few custom events, select a language, start a course, drop a course, and suddenly you have a dashboard where you can see your onboarding funnel and user churn. You build a complete app with security, real-time voice AI drivers, analytics, and state management. The method is to be a highly disciplined product manager and tell your AI what to pay attention to. You didn't write those boilerplate programs, but you controlled the architecture. You used higher-level architectural planning to replace the boring grammar work. To be honest, these tools we talked about today, whether it is Mainness helping you organize the messy download folder, letting you set up your own voice AI platform Hermes, synchronizing the AI heads of your mobile phone and desktop, or using pocket code file coding to create a complete application on the weekend. These tools are now mature enough. We no longer need to do the hard work. We just need to know how to manage these systems. If today's content is helpful to you, if you want to receive more latest and most up-to-date AI tool experience sharing and resources, you are also welcome to follow my IG, Face and Facebook search. You can find the AI column package. Feel free to help me leave a five-star review on Apple Podcast or your favorite platform. It can help me let more people receive the information on these AI tools. Let's become smarter together without working overtime. The recent development of AI is still so fast that it scares me. I don’t want to miss it and continue to listen to it tomorrow. AI man raps. My name is Tang Lanlan. We’ll see you tomorrow. Bye.

每天都有一堆新的 AI 工具冒出來是不是常常不知道該從哪裡開始? 別擔心,這裡是 AI 懶人報每天幫你精選五個最多人討論的 AI 工具影片 每天十分鐘,讓 AI 走進你的生活中,讓你不知不覺成為效率大師 今天要聊的東西我覺得真的很對我的胃 你有沒有想過,AI 不只是給你簡易幫你生一段程式嗎? 而是真的幫你動畫術,動鍵盤把事情做完 而且今天還有一個超實用的方法 可以讓你用 AI 在周末做出一個完整的應用程式 不用重頭寫每一行程式碼 我們直接開始 先講Manis 它最有感的地方是,它真的會動手幫你整理電腦 Manis 跟你平常用的那些雲端, AI 工具最大的不同 是它直接跑在你的電腦上 它有兩個很關鍵的權限,一個是可以存取你的本機檔案 另一個是,可以直接在你的作業系統上執行指令 它不是一個住在雲端的大腦,它是一個住在你電腦裡的工人 想像一下,你是一個接案子的人 或者就是一個電腦裡堆了一堆萬七八糟的普通人 你有一個資料夾裡面放了300個格式不一的發表 正常情況下,你可能會去問某個 AI 說幫我寫一個腳本來重新命名這些檔案 然後你打開中端機發現環境不對,搞了半天入境問題 最後弄完了,但你已經累了 用Manis 的話,你直接把資料夾指給它 說日期加客戶名稱的格式幫我重新命名所有檔案 它就直接做了 它會掃描每一個檔案,理解文件內容 然後在本機執行檔案系統的指令 幾秒鐘搞定 但真正讓我覺得有點嚇到的是 它在開發應用程式這件事上的能力 你可以叫它從零開始幫你做一個桌面應用程式 你描述你要什麼,它就寫程式碼 淘本機環境,除錯遇到的錯誤 然後把應用程式打包好 全部都在你的電腦上自動完成 以前可能要一個開發團隊花幾天才能搭起來的東西 它幾分鐘就搞定了 我個人最喜歡的功能是遠端控制 Manis 可以讓你用手機控制你的電腦 你在咖啡廳筆電開著放在家裡 突然想到有一個檔案要用 你打開手機上的應用程式 跟Manis說幫我在桌面找一下第三季的報告檔案傳給我 你的電腦就變成一個遠端控制的AI節點 我知道,把這麼大的權限交給AI聽起來有點嚇人 安全性的問題確實要注意 建議先在測試環境跑跑看 不要一開始就讓它碰到重要資料 但光是幫我把這些原始影片檔剪輯做縮圖整理到新資料夾 然後剩一個文字檔把連結都列出來 這樣一個指令就能從頭跑到尾 這種感覺真的很爽 回去手動點選會覺得非常痛苦 說到語音 AI下一個工具多寡解決的是另一個很麻煩的領域 如果你最近有在做語音AI的開發 你就知道那個Tone 表面上看起來很簡單 語音轉文字 丟給AI再把回應轉成語音 就這樣嘛 但真實的電話通話是一團混亂 使用者會打斷你 會突然沉默 會在AI正在講預算的時候問說 你是機器人嗎? 為了處理這些狀況 大部分人都用現成的平台 像VIP、Blend Retail 之類的 用起來很快 但帳單一來你就哭了 你在付大型語言模型的用量費 語音生成的費用、電信費用 還有平台自己抽的那一層 而且最麻煩的是 你根本不知道出了問題是哪裡的問題 那個系統對你來說是個黑盒子 客戶說完話後機器人聽了3秒 你根本不知道是語音辨識慢 模型回復慢 還是電話服務卡住 Dobra是一個開源的語音AI平台 專門給想要自己架設 自己控制的開發者用的 它支援Docker部署 你把程式碼抓下來 跑一個指令就進去了 但讓他聰明的地方 不只是開源這件事 它還有一個可以用拖拉的方式 把流程排出來的建構工具 你可以把整個對話流程畫出來 提示節點、資格審核步驟 呼叫外部服務更新客戶資料庫 分支轉給真人客服 對工程師來說 真正亮眼的是它的追蹤功能 你跑一個測試通話 比方說一個客戶打進來 詢問資格的機器人 在問公司規模和預算 Dobra會給你完整的對話逐字稿 延遲追蹤 每一個工具呼叫的記錄 還有實際的通話錄音 你不只知道機器人有沒有跑起來 你知道他為什麼這樣跑 或者當使用者突然打噴嚏的時候 他到底在哪裡出了問題 多管讓你自己帶大型語言模型 和語音合成的服務進來 他只提供語音引擎 流程建構工具 還有那些無聊但必要的分析和測試功能 這個專案目前還很早期 但它的方向是對的 在真正重要的地方 寫程式碼 用視覺化工具 處理對話流程 然後自己架設 這樣規模擴大的時候 才不會被平台的定價耗盡 自己掌控基礎設施 對正式產品來說很重要 但如果只是你自己個人的日常流程 你需要的是一個一直陪成你的助力 這就帶出了 Hermes 的最新更新 Hermes 的目標是 打造一個持久存在的自主 AI 可以在背景幫你監控系統 但他最近解決了 我覺得每天用 AI 最煩人的問題之一 記憶破碎的問題 想想你現在的設定是什麼樣子 你的 Cloud Code 住在你的桌子上 幫你做那些重度工作 他知道你所有的程式碼庫 你的寫程式習慣 然後你的手機助理可能是 跑在 Telegram 上的 Hermes 你在遛狗的時候把想法丟進去 問題是這兩個大腦從來不溝通 你在手機上傳了一個想法給 Hermes 但你做回桌機前 你還是要從頭跟 Core 解釋一遍整個概念 完全沒有交接 這次 Hermes 的更新 加了一個叫做 Agent Tye COS 的層 把 Hermes Telegram 機器人 直接連到你的 Cloud Code Core 那邊會追蹤你在本集做的所有事 監控你在不同模型上的用量費 記錄你的對話 甚至有一個功能是 他會在晚上做夢 也就是分析你當天的互動 然後建議流程改進的方向 Hermes 現在可以直接存取同一個記憶庫 他們還引入了一個叫做 Pantheon 的東西 這是一個在 Hermes 裡面的自定 AI 人格系統 每個人格都有指定的模型和角色 舉個例子 你想要做一個深度的行銷策略研究 但你不需要為了這件事 燒貴死人的定基模型用量費 你可以召喚一個人格 比方說叫他 Labyrinth 這個人格被設定成 使用比較便宜的模型 像是 Deep Seek 或者某個 Lama 的版本 專門用來做深度研究 你在 Telegram 傳訊息給 Hermes 交出 Labyrinth 幫我找五個針對屋頂公司的開發陌生客戶策略 Hermes 就把這個請求 路由到對的人格 用對的模型 然後把研究結果丟進你的 Obsidian 資料庫 對 他直接連 Obsidian 所以你在外面喝咖啡 用 Telegram 跟 Hermes 所取的 Power API 抓了十個本地屋頂公司的聯絡資訊 幫我起草開發信 Hermes 執行 API 呼叫 透過 CAP 的整合起草新建 然後把整個簡報存進你的 Obsidian 等你做回筆電前 Clarcode 已經有這些潛在客戶的完整背景 可以直接幫你做新應用程式的前段儀表版了 你從手機把無聊的研究和資料收集工作 委派給專門的 AI 人格 讓桌機上的 AI 負責執行 這就是終極的懶人物史設定 說到 Cloud Code 的 其實在聽了 Anthropic 內部工程師分享他們怎麼用它之後 我才發現大部分人用它的方式 根本是錯的 包括我自己 你一定有過這種時刻 你花了三十分鐘 精心打了一個超級完美的提示 定義了各種細節 你按下送出 AI 生出了還算不錯的東西 但在過兩個提示之後 他就完全忘記了最一開始的指令 你再用一個很脆弱的提示 來管理一個複雜的流程 這根本行不通 Anthropic 的工程師有四個規則 而且跟寫更長的提示完全沒有關係 第一個規則 提示技能包 不是提示 Cloud 軟體開發大部分都是重複性的工作 所以 Anthropic 工程師建立了 Cloud 技能包 一個技能包基本上就是一個資料夾 把某個任務的操作知識打包起來 你不用每次想起草一封信的時候 都重新解釋你的語氣和風格 你只要輸入一個指令 然後把原始文字丟進去就好 你開始用應用程式層的方式思考 BundCode 建立可以重複使用的小工具 第二個規則 一個技能包不只是一個提示放在資料夾裡 一個好的技能包有三個層次 第一層是描述 這是 Cloud 用來判斷他要不要使用這個技能包的東西 描述夠精確 Cloud 甚至可以在你沒有特別要求的情況下自動觸發他 第二層是指令 也就是一步一步的操作手冊 第三層是工具 也就是程式腳本 API 呼叫參考檔案 這裡才是真正有力量的地方 舉個例子 你想查一個網域名稱有沒有被註冊 你給這個技能包一個自定的拍放腳本 可以直接程式化的去查詢網域 這樣 Cloud 不是在猜 他是在執行程式嗎? 第三個規則 建立可以組合的小型技能包 不要建立一個大型的怪獸 我的直覺是建立一個超大技能包 一次搞定靈感發想 腳本撰寫和社群文案 Anthropic 的團隊說 這是個陷阱 你想改Lintint帖文的格式 你去編輯那個巨大的技能包 然後你的YouTube腳本突然壞掉了 他們的做法是建立小的聚焦的技能包 因為Cloud可以自動在他們之間協調 哪個技能包壞掉了 你馬上知道問題在哪裡 還有一個很聰明的優化方式 把腳本存在技能包裡面 程式碼是確定性的 AI 不是 如果Cloud寫了一個Python腳本來格式化投影片 不要每次都叫Cloud重新寫一遍 把腳本存進技能包的工具成 下次Cloud直接執行那個腳本就好 你用便宜的確定性的本機運算 換掉了貴的不確定的AI用量費 第四個規則 提示應該要隨著每次使用變得更聰明 你在聊天界面打的普通提示 視窗一貫就消失了 用技能包的話 技能包會一直留著 Anthropic 的工程師 把技能包當成會負利增長的資產 如果Cloud犯了錯 他們會問Cloud 回顧一下我們剛才的來回 更新你自己的技能包指令 讓你以後不要再犯同樣的錯 用了30天的Cloud 應該要比第一天好非常多 因為每一個邊緣案例和個人偏好 都已經永久寫進技能包的架構力了 這種建立模組化可預測系統的思維 帶出了今天最後一個主題 如何真正用AI把應用程式做出來 今年初5coding這個詞流行起來就是那種 你描述你要什麼 讓AI去做 自己不看程式碼的方式 如果你在超過玩具等級的專案上試過這個 你就知道結果是什麼 前三個功能感覺像魔法 你飛起來了 到了第五個功能 你叫AI把一個元素置中 它不知為何把你的資料庫結構搞壞了 現在你面對著幾千行你沒有寫的程式碼 在一個你完全不理解的架構上 速度沒有方向就是一場災難 最好的平衡點是開發者社群 現在說的practical vibe coding 核心是你專案跟目錄裡的一個檔案 agentsmd 像cursor,windsword,clockcode這些工具 匯讀你的專案檔案來取得背景 如果你沒有一個中央規則手冊 你就要在每一個提示裡 重新打你的技術站和資料夾規則 有一次你忘了說你用sustent做狀態管理 AI就自己裝了redux 然後你的程式碼庫突然有了分裂人格 agentsmd解決這個問題 他從一個角色開始 你是一個資深的react native工程師 你把清晰度放在抽象化之前 就這一句話 就成為每一行生成程式碼的過濾器 接著你定義技術站 export native1做樣式 cursor做身份驗證 sustent做狀態管理 每個一行 這把AI所死了 如果AI一直犯同一個錯誤 你在agentsmd加一條規則 例外 safe area view使用標準樣式 其他鎖的地方使用native1 class 你解決一次 記錄下來 ai就不會再犯同樣的錯 agentsmd鎖定之後 你進入提示階段 pract go by coding的提示結構 每個提示有四個部分 第一部分 先讀agentsmd並嚴格遵守 這強迫上下文視窗 重新再入規則 第二部分 一個單一任務 不是做好新手引導 設定身份驗證 然後錄有到首頁 就只是實作新手引導畫面的介面 第三部分 行為限制 你告訴 AI他不被允許碰那些東西 不要更動現有的驗證流程 不要在應用程式裡 曝露機密諮詢 你在保護你已經可以跑的程式碼 第四部分 設計參考 貼一張截圖進去 AI馬上理解視覺版面 省掉你解釋三段間距和編距的時間 用這個框架 以身份驗證為例 你不想要AI自己發明一個自訂的安全始作 你要探用Curk 你把最新的CurkExpo文件存成markdown檔案 貼近你的提示說 用真實的Curk身份驗證 取代模擬的驗證流程 10座電子有9OTP和Google Auth 成功後導向首頁使用提供的文件 因為AI讀了你的Agent3D 它知道要用你現有的Native Y的樣式 因為你提供了最新的文件 它不會用訓練資料理經過期的API 幾分鐘內你就有了正式等級的安全身份驗證 真正考驗這個流程的是整合複雜的即時功能 像是語音、AI家教 目標是要有一個即時低延遲的AI語言家教對話 你用Stream的React Native SDK做音訊傳輸 然後連到一個AI老師 把它想成一位會即時聽你講話 根據課程目標回應的AI 你不想把機密經料曝露在應用程式裡 所以你用Expo後端路徑做安全的用量費提供者 應用程式向伺服器要求通行 伺服器生成通行 應用程式加入通話 你提示AI使做Street明訊通話設定 用Expo後端路徑做通行生成 不要曝露機密 在介面加入載入中以加入和錯誤狀態 音訊傳輸跑起來之後 你再啟動一個獨立的Python服務座 AI老師你告訴AI 建立一個Python服務 用OpenAI即時功能做為AI語言老師 他必須從通話資料讀取選擇的語言和課程目標 突然間你打開手機上的應用程式 點一個西班牙語問候的課程 就在跟一個知道你應該學哪些單字的 AI家教進行即時的語音對話 如果AI老師因為顏色問題一直搶著說話 你寫一個針對性的修正提示 移除相機按鈕 建立一個至中的按住說話麥克風按鈕 當我按住按鈕 立刻讓AI靜音也要聽我說話 使用體驗瞬間就像一個做得很完整 真的像正式產品的應用程式 最後你把Posthawk加進去做分析 用他們的AI精靈他掃描你的Reacognitive程式碼庫 自動把你的應用程式包在一個分析提供應用程式裡 然後開始追蹤畫面瀏覽 你加幾個自訂事件 選擇語言 開始課程 放棄課程 突然你就有了一個儀表板 可以看到你的新手引導漏斗和使用者在那裡流失 你建了一個有安全機制 即時語音 AI驅動 有分析 有狀態管理的完整應用程式 方法是當一個高度有紀律的產品經理 告訴你的AI注意要做什麼 你沒有寫那些樣板程式嘛 但你控制了架構 你用比較上層的架構規劃 換掉了那些無聊的語法工作 說真的今天聊的這些工具 不管是mainness幫你整理亂七八糟的下載資料夾 多寡讓你自己架設語音 AI平台Hermes 把你的手機和桌機的AI腦袋同步起來 還是用pocket code file coding的方式 在周末做出一個完整的應用程式 這些工具現在已經夠成熟了 我們不用再做那些苦力了 我們只需要知道怎麼管理這些系統 如果今天的內容對你有幫助 想收到更多最新最即時的 AI工具心得分享跟資源 也歡迎追蹤我的IG、Face和Facebook搜尋 AI欄語包就找得到 隨手幫我在Apple Podcast 或你常用的平台留個五星好評 可以幫助我讓更多人 一起接收到這些使用的AI工具資訊 一起變聰明不加班 最近的AI發展依然快的讓我害怕 不想錯過明天再繼續收聽 AI男人暴羅 我是湯嵐嵐 我們明天見 拜拜