13分鐘搞懂全網爆火的循環工程 Loop Engineering！

If you've recently come across terms like "Loop Engineering" but have no idea what they mean, congratulations— you've clicked on the right video. " Chain Engineering" was all the rage not long ago, and now "Loop Engineering" is here. It makes me seriously suspect their KPIs include a requirement to invent a new term ending in " Engineering" every quarter. But this time, after looking into it carefully, I found that it's quite different from what came before. Boris, the head of CloudCode, says he no longer sends prompts to the cloud; his job now is to write loops. Peter, founder of OpenCloud, put it even more bluntly: " You shouldn't be sending prompts to AI coding assistants anymore; you should be designing the loops that send the prompts for you." Later, even Andrew Ng wrote a dedicated column about this, so I’ve scoured the entire web for everything related to "Loop Engineering ." Today, we’re going to use a timeline to clarify these four terms once and for all: "Prompt Engineering," "Context Engineering,""AI Workflow Engineering," and "Loop Engineering." I’ll explain them very simply, so you don’t need any technical background to understand. However, I have to be a bit of a buzzkill and admit that after going through all the material, I found that to this day, no one has actually provided a real-world use case where " Loop Engineering" works wonders—not a single one. So, why should we spend over ten minutes trying to understand it? I’ll give you the answer later. Without further ado, let's get started. Before diving into the details, let me share the three most important conclusions: first, these four terms aren't four separate things; they are four different stages of the same thing . Second, there is an underlying pattern behind these four terms. Once you understand this pattern, you’ll be able to predict what the next buzzword in AI will be yourself. I will explain this pattern halfway through. Third, the core of Loop Engineering comes down to one single point: removing the person giving instructions to the AI from the process—and that person is you. In the future, we won't be sending instructions; we will be designing the system that sends instructions for us. Keep this sentence in mind, because everything that follows will revolve around it. Alright , let’s start from the beginning. On November 30, 2022, GPT was first introduced to the public. For the next two years, the whole world focused on the same thing: how to speak to it effectively. Andrej Karpathy famously remarked that the hottest new programming language is prompting, also known as Prompt Engineering. By June 2025, Shopify CEO Tobi Lütke mentioned on social media that he preferred the term "Context Engineering." He argued that in large language models, it describes our needs more accurately than prompt engineering. Karpathy agreed, and thus, "Context Engineering" became the standard. Moving to February 2026, Mel, the creator of Transformer, proposed an idea: whenever AI makes a mistake, instead of blaming the AI, we should modify the system so the error becomes structurally impossible. This led to AI Environment Engineering, followed by June 2026, which brings us to today's topic: Loop Engineering. Next, we will break down these four stages one by one. Let's start with the innermost layer: Prompt Engineering. Everyone is familiar with this—you tell the AI, "You are a professional customer service agent, please be kind to my clients," and then ask it a question; it will respond in a professional tone. Simply put, prompt engineering manages that single conversation. To put it more plainly, think of it like the "special instructions" on a food delivery order. You ask for less spice, no cilantro, no utensils, and the chef follows your lead. The clearer your instructions, the closer the meal will be to what you wanted. But here's the catch: even the best instructions have a limit. If the kitchen doesn't have the ingredients, no matter how well you write your request, it can't create the dish. For example, if you ask the AI, "How many cups of bubble tea can fit in the distance between my home and the office ?" That type of question can be answered with just a prompt, because it doesn't require any external information. But if you ask for this week's company work schedule, it's stumped because it has no way of accessing that information. This is where the second layer, Context Engineering, comes in. The principle is very simple: your prompt is stored within the AI's context window. You can think of the Context Window as the AI's short-term memory, or perhaps its workbench. So, how much space does a single prompt really take up? It certainly won't fill up the entire workbench. So, people simply gave it more authority, letting it fill up this workbench itself. This allows the AI to read your files, edit your documents, and even connect to your databases and various external applications. You can think of MCP as a universal plug that lets the AI connect to different pieces of software. This way, when you ask the AI about the latest news, it can search the web, organize the info, and then give you an answer. So, to summarize context engineering in one sentence: it's letting the AI find the information it needs on its own. Using the food delivery example again, prompt engineering is like you writing exactly what you want in the notes, while context engineering is like giving the chef access to the fridge and procurement system so they can source the ingredients themselves. Sounds pretty perfect, right? But it has a flaw, and it's a flaw you've likely encountered before. Have you ever had the experience of chatting with GPT or Claude about a long topic, only for it to start forgetting things, repeating itself, or giving you contradictory answers, making you think, "Wait, didn't we agree on that?" The reason lies right here. Context engineering doesn't have a fundamental defect; it just has limitations. Because it can't handle long-term tasks, as those require an amount of information that far exceeds what the workbench can hold . The AI's solution is to summarize whenever it's nearly full, compressing the old information into a short paragraph. It sounds clever, but every time it summarizes, some details get lost. After five or ten times, the most important things from the beginning are gone. You can imagine it as a new intern: you gave them full system access, and they are diligent about looking everything up themselves, but they have a fatal flaw—they don't have a notebook and rely entirely on their memory. You give them three tasks in the morning, but by the afternoon, they only remember two. What do you do? The answer isn't to force them to try harder to remember, but to provide them with the right office environment. And this is the third layer: AI Harness Engineering. It no longer manages just what's inside the AI's "brain," but the entire environment surrounding the AI. It helps the AI manage context externally, breaking down large requirements into a series of stable, executable tasks. Back to that intern analogy: what we need to do now is give them a dedicated desk, a manual, and an access badge. Then, set a rule: any mistake made once must be documented in the manual, and that same mistake is never allowed to happen again. That’s exactly what Mel meant. When AI makes a mistake, don’t scold it; we should be modifying the AI’s environment instead. Let’s compare the three scenarios we just discussed. If you ask ChatGPT, "How many cups of milk tea can fit between my home and office," that’s prompt engineering; it can figure that out on its own. If you ask GPT, "What’s the big AI news this week," that’s context engineering; it has to search the web to answer you. But if you open Claude and say, "Clone this entire website from top to bottom, " you need AI environment engineering, because a full website has hundreds of files. Relying solely on context engineering would cause it to lose track halfway through and just get stuck. AI environment engineering provides an external mechanism to manage the context and runtime, allowing the AI to complete a massive volume of tasks in one go. By now, I believe you should have a vague sense of something. Have you noticed that all three layers hide the same thing: a loop? In context engineering, the AI calls tools one after another, only stopping once it feels it has sufficient information—that’s a loop. In AI environment engineering, the AI keeps a task list outside the context window, working through it one item at a time, checking them off as it goes—that’s another loop. To put it simply, we’ve been doing the same thing all along: layering one loop on top of another. There is a pattern behind the four terms we mentioned at the start of the video. Look at the scope each one manages: prompt engineering handles a single paragraph, context engineering manages a desktop, and AI environment engineering manages an entire office. Doesn't that expand layer by layer? Loop engineering is essentially wrapping another loop around AI environment engineering to drive the entire environment from the outside. So, what exactly does it manage? Let me point out a fact you might not have noticed: in every example we’ve used, whether it’s milk tea, AI news, or cloning a website , they all share one commonality—you are the one asking the AI what to do. The instructions always come from you; if you don't speak, the AI just waits there. Loop engineering, on the other hand, aims to delegate the act of prompting itself, letting the AI decide what to do and issue its own instructions. It isn't managing the AI, but you; the link it intends to replace is you sitting at the computer constantly typing. So, prompt engineering focuses on the statement itself, context engineering looks at the environment surrounding that statement, while loop engineering is concerned with the operational rhythm of the entire system. What exactly does this loop contain? There are six actions in total: receiving the task, executing the work, checking the results, processing feedback, adjusting the plan, and doing it again, looping continuously until the goal is achieved . Returning to the practical example, this stage is where the AI officially becomes a full-time employee. You don't need to assign tasks every morning; it checks the schedule, finds work on its own, inspects its output, and fixes any issues. You only need to review the work the next morning. Now, let me present a concrete example to show you the full picture. Suppose I built a website to track the RMB to AUD exchange rate and list the fees for major remittance platforms. I asked CloudCode to build it, and it used prompt, context, and AI environment engineering to finish the job. Everything went smoothly until the maintenance phase, because exchange rates and fees change daily. To keep the site useful, I had to ask the AI to update it every day, and after launch, users reported various bugs that I had to ask the AI to fix one by one. Do you see it? I had become the bottleneck of the entire process. Let's change our approach: I set a scheduled task in the cloud to have it check for new data every hour and update if necessary. For bugs, I let it read user feedback and resolve them independently. By resolving them, the entire situation changes completely. I am no longer the one chasing the work; this loop has started running on its own. Furthermore , I have already installed skills and plugins in the Cloud environment, so the AI can read existing knowledge bases directly. It can supplement information as it works, call on various sub-agents to check its performance, and handle multiple fix tasks simultaneously. Alright, let's return to the question we left at the beginning. To this day, no one in the market has been able to showcase a case study where loop engineering has worked wonders, so many people feel this term is just a gimmick. However, after reviewing all the materials, my conclusion is that it’s fifty-fifty. The "gimmick" part is the term itself, as it remains highly theoretical and mostly related to programming, making the barrier to entry quite high for the average person today. But the other half is the direction it points to. In his column, Wenda didn't just point to the code level; he expanded outward and proposed three loops. The first loop is AI writing, testing, and revising on its own until it meets the standard. The second loop is you adjusting your requirements based on what it produces. The third loop is letting real users provide feedback after using it, so we can turn back and adjust our direction. So, the innermost first loop can be left to AI, but the two outer loops still belong to humans. Therefore, the point of this is never the AI itself, but rather you transforming from someone who is constantly typing into someone who sets the direction, sets the standards, and performs the final acceptance. So, what can we ordinary people do today? There are three things in total, and let’s start with the simplest one. First: write down your routine tasks. See what you do repeatedly every week and what your internal criteria are when you do them. This document will become your future skill set. The first step of loop engineering isn't technology; it's being able to clearly articulate what you are actually doing. Second: start with acceptance work, not with full automation. Many AI tools you use now have scheduling features. Set a low-risk one, like having it summarize industry news for you every morning. You spend two minutes scanning it, and after a month, you’ll know where it’s reliable and where it isn't. Third: always keep an AI for inspection , and remember, it must not be the same AI that does the actual work. You can open a new chat, paste in your standards, and let it specifically pick out faults. This you can use today, and it will be effective immediately. Finally, let's summarize the core content of today in one sentence: prompt engineering manages a single paragraph, context engineering manages a workbench, AI environment engineering manages an office, and loop engineering manages you. Also, I previously made a complete video on AI Agents that explains their underlying logic very clearly; if you're interested, you can click here to watch it. OK, that's all for this video. If you found it helpful , please like and subscribe to the channel. See you in the next one, bye!