← Back to search

EP68|Save Money in Practice! Essential Skills During the AI Agent Boom: Hermes Updates and Circular Engineering to Help You Earn Passive Income!

AI熱搜報 · 2026-06-25 · 13 min
relevance 89 99 words Episode page ↗ Audio ↗
Show full episode description
喵~我是DaeDae,一隻幫你追AI熱搜的貓!AI熱搜報,幫你劃重點!最近 24 小時的 AI 圈簡直是 Agent(代理人)的實戰競技場!從底層邏輯到高 CP 值的省錢組合,甚至連「提示詞工程」都進化成了「循環工程」。DaeDae 幫各位鏟屎官整理好了最強自動化攻略,讓你的 AI 不只會聊天,還能幫你動手幹活,一起輕鬆摸魚賺罐罐! 【本集肉球筆記】 1️⃣ 掌握 AI 代理人核心術語,溝通不再鴨子聽雷 👉 深入淺出拆解現代 AI Agent 的五大底層邏輯,讓你精準指揮代理人。 🎯 重點包含:`agents.md`(代理人說明檔)、`agent skills`(專屬技能)、`MCP`(模型上下文協議)、`agent-to-agent communication`(代理間溝通)與 `sub-agents`(子代理人)。 🔗 https://www.youtube.com/watch?v=k5jYwyhDMxA 2️⃣ 極致高 CP 值玩法:Hermes Agent 搭配 Minimax M3 👉 透過「租用便宜大腦」策略,打造效能最強且最省錢的自動化工作流。 🎯 核心在於「動態模型路由(Dynamic Model Routing)」,利用 `Firecrawl` 抓取資料並搭配 `Minimax M3` 模型處理,大幅降低自動化任務的營運成本。 🔗 https://www.youtube.com/watch?v=xxpc_6GAgGo 3️⃣ Hermes Agent 重磅更新!桌面變成自動化指揮中心 👉 代理人功能大躍進,原生整合通訊軟體,讓自動化無縫切入生活場景。 🎯 重點功能:`Native iMessage` 與 Telegram 原生整合、`Background agents`(背景執行)、`Unreal MCP` 擴充、`Profile builder` 以及 `Skills Hub` 技能庫應用。 🔗 https://www.youtube.com/watch?v=bQ1LCFrwj08 4️⃣ 誰說提示詞工程已死?最強四步驟「循環工程」大揭秘 👉 教你如何設計自動化迴圈,讓 AI 自我修正產出,實現高品質結果。 🎯 透過 `Loop Engineering(循環工程)` 的四步驟工作流,取代單次提示,利用迭代修復技術提升複雜任務的成功率,真正達成躺平賺罐罐。 🔗 https://www.youtube.com/watch?v=JirDfgJcJFU 5️⃣ 免費無浮水印!最強 AI 影片生成器 Grow Cup AI 教學 👉 手把手帶你產出高品質病毒式影片,是鏟屎官省錢做內容的秘密武器。 🎯 涵蓋 `OI OI AI` 帳號設定、文字轉影片、`Aspect Ratio`(長寬比)調整、畫質控制及領取每日免費點數的隱藏訣竅。 🔗 https://www.youtube.com/watch?v=lGc1_ul-s8w 想摸魚跟上 AI,記得訂閱 AI熱搜報!我要去睡午覺了,我們明天見🐾 Powered by Firstory Hosting
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Make AI agents useful and affordable through explicit rules, model routing, reusable workflows, and measurable verification.
Benefits
  • Reserve expensive models for difficult judgments
  • Clean website content before sending it to models
  • Delegate research into clearly scoped subtasks
  • Record workflow state between repeated runs
  • Test creative concepts before committing a production budget
Use cases
  • Marketing example: delegate competitor price checks, feature comparisons, and review collection before compiling a summary.
  • Suggested Hermes and MiniMax M3 workflow: use stronger models for strategy and cheaper models for website organization and video summaries.
  • Meeting-preparation example: request research from a phone, then resume the work on desktop.
  • Daily AI news workflow: collect stories, draft a roundup, verify sources, and refine openings using previous performance.
  • Suggested product-ad example: script Mother's Day videos for handmade desserts and test different creative versions.
Tools / build
0:00 / 0:00
🌐 This transcript was automatically translated to English from the original.
Ah, I'm DaeDae, a cat that helps you track AI trending topics. The AI Trending Report helps you highlight the key points. Do you feel that AI is developing too quickly, with new things to learn every day? Don't worry, DaeDae is here to help. Today, we've selected five of the hottest AI tool updates. With just 10 minutes a day, you can easily keep up with the AI wave and become the smartest "slacking off" master. Before diving into today's five major tool updates, DaeDae wants to share a little life insight with everyone. Recently, many Zen Food Officials have complained to me, feeling that if they don't quickly learn all the AI tools, they seem to be on the verge of being abandoned by the world. Actually, everyone really doesn't need to be so anxious. AI is like a newly purchased automatic feeder or robot vacuum at home. Its purpose is to give us more time to sunbathe and take a nap, rather than turning us into servants who run after machines all day. Change your mindset and treat AI as a little assistant that represents efficiency. You'll find that exploring these new gadgets is quite stress-relieving. If you're not sure, it can also help you save a lot of overtime hours. Today's episode is like a practical, no-nonsense cheat sheet. It's not the kind that just talks concepts and leaves you still foggy after listening. Instead, it walks you through five directions that are genuinely relevant to daily work. The key point is actually just one: AI isn't just about chatting. The real value lies in how you set it up, connect it, and deploy it in the right places. Only then will it truly help you do work—or even help you earn money. And if you only want to remember one thing today, it's this: the most useful thing about AI right now isn't how beautifully it answers, but that it's starting to be able to complete tasks step-by-step according to a process. Before, many people thought of AI like a very good conversational colleague—you ask it, it replies, and then that's it. Now, what people are chasing is making it not just reply, but be able to do things step-by-step on its own, check and correct itself, and then hand back the results. Grab this main thread, and the next five topics will be much easier to follow without getting distracted halfway through. First, let's talk about five key terms related to AI agent assistants. Many people keep hearing about "AI agent assistants," but when they actually try to use them for work, they start getting confused. Why do some seem very powerful while others are often unstable? In plain language, a truly useful AI agent isn't just about the model itself. It also needs rules, tool connections, and division of labor to work reliably. First up is Agent Sample. You can think of this as the work rules an AI must follow before it starts. For example, which files cannot be changed, what format responses should take, and whether it should ask a human before tackling a problem. Next is Agent Skills. This is more like a skill set that can be pulled out when needed, rather than having everything turned on all the time. Then there's NCP. You can think of this as a plug that connects the AI to external tools, allowing it to access databases like Notion, Stripe, or various service interfaces. Without it, the AI often just acts blindly. A2A is what allows different AI agents to pass tasks to each other and divide the work, rather than forcing a single AI to play all the roles at once. Finally, sub-agents are large tasks broken down into smaller chunks so several small helpers can work together, with results being collected and returned. In a corporate work context, this is easy to understand. For example, in marketing, one AI agent first follows the work rules to organize competitor data, then uses tool connections to scrape website information and data. Next, other AI agents are assigned to check prices, compare features, and compile reviews. Finally, a clean summary is handed back. What you see isn't just a paragraph of answers, but a full, actual workflow of doing things. Freelancers can do the same thing. For example, before submitting a proposal, read brand materials, gather market information, and compile a first draft. The end result isn't starting from scratch, but starting from a draft that can be edited. To summarize: a smart cat is great. But if it also understands house rules, knows how to open doors, calls its companions, and can even send little cats to run errands, then it's truly capable. DaeDae helped highlight the key point: the real value of an AI agent isn't being better at talking, but being better at following rules to complete tasks. Okay, let's move on to the next tool. The second one is Hermes Agent paired with MiniMax M3. The point isn't which model is the strongest, but that you shouldn't blindly throw everything at the same expensive brain. Many people just use Code or ChatGPT for everything all the time. The result is that even for simple tasks like scraping a webpage to look at images or watching human videos, they're still burning money on high-cost models. The benefit of Hermes Agent is that it can preserve the entire workflow and then swap out different model "brains" depending on the task. MiniMax M3 here is a cheap and flexible option. It can handle very long content at once, process text, images, and video simultaneously, and even do web browsing. For many people who don't need top-tier specs every single day, this kind of pairing is actually very practical. Just remember one concept here: not every job needs the general manager to come out. Models are suitable for tasks that are difficult to judge as high-risk and require creativity; cheaper models are better for organizing, converting, and extracting information, which is often sufficient. Pairing this with FireCrawl, a tool that specifically extracts the truly useful text from websites, avoids dumping large blocks of web pages, raw code, and noise into prompts. Cleaning the data first typically yields more accurate results at a lower cost. In practice, small studios or project managers planning product limits can first use stronger models for strategy, then delegate repetitive tasks such as website organization, image analysis, and video summary extraction to MiniMax M3. This approach saves money and increases speed, while reducing the tendency to abandon work halfway through. The cat analogy summary is simple: if your cat uses only one toy for everything, your home will inevitably become messy and expensive; a cat that knows how to choose the right toys is truly living well. Model routing is not random; it involves first distinguishing which tasks require deep thinking and which do not. The third point concerns the latest update to Hermes Agent. This is not a single feature overhaul but a transformation making it feel more like a real assistant. Unlike previous versions that appeared powerful but were actually cumbersome to use, many AI agents are not unimpressive but are simply difficult to use daily. This time, Hermes Agent added support for iMessage via Futon. In plain terms, you can now use your phone messages to have it check data, open browsers, and process files, without needing to return to your computer to give commands. There is also a background assistant mode where routine tasks run autonomously in the background while you continue to issue commands and monitor progress at the front end. You are no longer stuck waiting for a single task to complete. Additionally, the desktop interface now supports multiple chat windows, built-in model switching, and a built-in terminal, and it supports Enreal Engine 5.8 NCP. This means it can now handle more serious 3D game production, not just document work. Even if you are not a game developer, the key takeaway is that it is no longer just a text toy but is moving toward a more complete work platform. In practical application, you can allocate one Hermes instance for setting up configurations and writing code, another for research, and another for organization. Use your phone for ad-hoc requests, and let routine tasks run in the background gradually. For example, you can ask it via phone in the morning to organize materials to review before a meeting; by the time you arrive at the office, it will have completed half the work, and you can simply take over when you open the desktop. This seamless handoff is the key to whether any tool can truly become part of your daily workflow. The cat's conclusion is that this is not just about making cats smarter, but also giving them access to messaging software, physical boards, and better memory. It upgrades them from cute chaos to organized chaos. The real value-add is moving AI agents from occasional play to daily use. If that didn't convince you, watch this V-Day video and you'll be awake. The fourth topic is cyclic workflow design. Another important concept is that prompt engineering hasn't died. Now, people often say prompt engineering is outdated, and the future is cyclic workflow design. But that's not quite accurate. A more precise way to put it is: prompt engineering is the foundation; it's just that people are now building entire workflows on top of it. In plain terms, cyclic workflow design turns prompts into repeatable, verifiable, and recorded work processes. The core four steps are simple: Trigger, Execute, Verify, and Status/Record. Trigger is when to start the process. Execute is telling the AI to do something. Verify is defining what success looks like first. Status/Record is logging what was done in the previous round so the next iteration doesn't start over. Tools like Coco or Codex can handle the execution. Tasks that truly fit this workflow are those with quantifiable outcomes, such as making Python code run faster or increasing headline click-through rates. But if you just say "write better social media posts," that's too vague. What does "better" mean? Who wins? These need to be clarified first. Take a content creator example: suppose you need to publish an AI news roundup every day. Previously, you might have manually searched for information and wrote manually daily. Sometimes you'd get inspired and do it, other days you'd just be lazy. With cyclic workflow design, you can make it a fixed routine: grab the news, create a first draft, check for missing sources, then review previous days' performance to refine the opening and closing of the post. After running this for a while, it will gradually align with your work rhythm. The cat's summary is: a cat chasing the same laser dot isn't a strategy. But if it remembers where the laser dot was yesterday, you won't fool it as easily. Many people aren't bad at using AI; they're trying to go fully automated too early, without first defining success criteria. The final point is about OIAI (also called Growcar PI), a free AI video workflow focused on no watermarks and daily free credits. This pain point resonates strongly with many creators. You want to quickly make animated stories, product ads, or short videos, but most video tools are either too expensive or charge for features right from the start. Plus, large watermarks on the final product are really annoying. OIAI is special because it wraps the entire production process into a guided interface. You can start with Story AI, a script tool, to write the script, define characters, create scene storyboards, add sound, and finally assemble the video. You can also use Character Designer for characters, Scene Designer for scenes, Rodact Test for product ads, or Free Creation for generative short videos. For beginners, this approach is very helpful because it sets the order for you first, then pairs it with ChatGPT to help write prompts aligned with OIAI's candidate styles and video models. This allows you to create multi-angle videos with dialogue, and download the watermark-free version directly. In practice, small businesses can use this to quickly create social media product ads. For example, if you're selling handmade desserts and want a Mother's Day theme, you can first use Story AI to generate a short script, then use Project Test to create different versions and see which one grabs attention. New creators can also use daily free credits to test character-based shorts without spending a full editing budget on day one. The cat's summary is: this is like someone directly I'm giving you a free studio, plus a little snack every day, so that your cat, who's always acting out, can finally produce the work it believes is its destined masterpiece. However, while these free points-based platforms are great, they're usually best suited for testing the waters. If you're planning long-term, high-volume output, you should evaluate stability and consider paid plans, which are more practical. Anyway, today's five topics may seem very different, but they all talk about the same thing: AI truly becomes useful not because it suddenly becomes super-omniscient, but because it increasingly aligns with your daily habits, budget, and work content. You don't need to master everything; just find the direction that best suits you. Maybe it's setting AI agents with more structure, maybe it's learning model routing, maybe it's turning a fixed task into a verifiable loop process, or maybe it's using free video tools to produce content. Any of these could be the step that makes you more efficient at "slacking off" today. If you're still a bit confused after listening, let me recap for you: First, give AI agents rules, tools, and division of labor—they shouldn't just reply to messages. Second, learn model routing—don't use the most expensive model for everything. Third, let AI assistants truly integrate into your daily life: you can call them on your phone, let them run in the background, and use different settings for different tasks. Fourth, upgrade your prompts into repeatable, verifiable, recordable loop processes. Fifth, use free AI video tools to get content production started at low cost. If you want to stay updated on AI, remember to subscribe to AI Hot Search Report. If today's content gave you something valuable, please leave me a five-star review on Apple Podcasts. Search for "AI Hot Search Report" on Facebook, Instagram, and Threads to find me. That's all for today's AI Hot Search Report. I'm going to take a nap. See you tomorrow. Bye-bye, kitty!