Is the era of "loops" next after prompts? A simple four-stage guide to the evolution of how we communicate with AI.

本ページはプロモーション（広告）を含みます。

🧭 Vol.1 "People who are good at prompting" are no longer the main characters

Until recently, being "good at prompting" was a compliment. It referred to people who could give precise instructions and elicit the desired answers from AI. Techniques like specifying roles, such as "You are an accounting expert," were also frequently introduced. I myself have actually used them quite a lot. However, since the beginning of 2026, voices have started to be heard from the forefront of AI development saying, "Let's stop writing prompts."

Wait, what do you mean we're stopping writing prompts? I've introduced a lot of role-assignment techniques in my English learning articles. I felt like just writing "You are an English teacher" made a huge difference in the corrections. I'm a little shocked that it's no longer the main focus. I'm sure there are people who read my articles and used it. So, what will be the main focus from now on?

In development environments, the focus is shifting from individuals giving individual, clear instructions to those creating systems that continuously provide instructions. This is sometimes referred to as a shift from "Human in the Loop" to "Human on the Loop." It's a change from people working within a loop, taking on tasks one by one, to standing outside the loop and overseeing the whole picture. This evolution can be organized into four stages: prompt, context, harness, and loop.

It's fascinating how simply changing "in" to "on" completely alters how we interact with AI. Even with English learning, honestly, typing prompts like "correct this essay" every day was a bit of a hassle. It would be amazing if I could create a small learning loop that remembers my weaknesses and even creates practice menus for me. That would make everyday life so much easier. But we need to start with the basics, right? What exactly do you mean by "improving prompts"?

✍️ Vol.2 Prompt Engineering

Prompt engineering is the stage where you refine the very wording of each instruction you send to the AI. Even when asking for the same thing, the quality of the answer can vary greatly depending on how you ask. For example, asking "Please proofread this English sentence" is less effective than asking "Please rewrite this English sentence to make it sound more natural as a business email." You could call it refining the "way" you ask the AI. This is the simplest form of this process.

I do that almost every day. Just adding a brief statement of purpose makes a huge difference in the tone of the English response. Like, "casual," or "suitable for email." But, based on what we were just discussing, specifying someone's role, like "You are an expert in...", is outdated, right? It's often in introductory books and seems effective, but why isn't it enough?

Simply assigning a role to an AI doesn't tell it "who" wants the answer and "why." What's important is the purpose, goal, and conditions, rather than the title. For example, just adding a phrase like "I'm a beginner in accounting" will encourage it to provide an explanation that avoids technical jargon. Adding a goal like "in a format that can be used immediately in practice" and conditions like "in informal language" will make it even more accurate. The key is to convey the situation, not just the role.

It's more important to explain what you'll be using it for than your title, right? If it's for learning English, I guess I should write something like, "I want to acquire business English skills so I won't be embarrassed in real-world situations." That certainly seems like it would be more tailored to my situation. If my purpose is clear, my own study focus will probably stay consistent. But it's a pain to explain my background and weaknesses from scratch every time. I wish the AI could remember what we've talked about before and the mistakes I often make.

🗂️ Vol.3 Context Engineering

That's precisely where the second stage of context engineering comes in. It involves designing the entire set of information the AI looks at when it provides an answer, including the conversation history, the documents provided, and past interactions. There's an upper limit to how much information the AI can see at once, and this is called the context window. It's said to be able to hold about 10 volumes of books, but it's not infinite. Designing what to put in it is crucial. And that prompt we saw earlier is just a small part of that big box.

It's not about "what to ask for," but "what to show them." When talking to people, the way the conversation goes is completely different depending on whether you share the same assumptions. I understand that even if you write a good prompt, it will be off the mark if the AI doesn't know the background. It's interesting that the prompt is part of the context. Even with English conversation teachers, the more they know about me, the better they teach. So, what kind of information should I specifically show the AI when learning English?

For example, you can provide a list of expressions you've frequently gotten wrong in the past, or your learning history, such as "I often make mistakes with tenses" or "I stumbled on this expression last week." Just doing that will transform the corrections from general feedback to personalized feedback. They'll also notice if you're repeating the same mistakes. However, more isn't always better. If you include irrelevant information, you'll end up with irrelevant answers, so the trick is to focus on recent mistakes or current topics.

So, instead of showing everything, you only show what's necessary. If it were me, I'd probably just show the mistakes with articles and prepositions that I'm working on this week. That alone would make the corrections much more tailored to me. The trick is not to be greedy and give everything away. But what if the AI gives a plausible-sounding but incorrect correction? It's a little scary because when it says it with such confidence, you might end up believing it at face value.

🦺 Vol.4 Harness Engineering

This is where the third stage, harness engineering, comes in. A harness is a safety device used in mountaineering. Asking "Please write a test" is a prompt, but the harness is the mechanism itself, which dictates that "you cannot proceed to the next step unless the test is passed." It's an external agreement for the AI. For scheduling, if there's a mechanism that "rejects attempts to schedule appointments after 5 PM," you can confidently leave it to the AI.

It's not about writing the instructions in the code, but more like creating a safety fence around the AI. The secretary analogy is very clear. If it's decided from the start that the AI can't enter after 5 PM, it won't be a problem even if the AI moves around a bit freely. You communicate the same rules to a human secretary, right? But even if you say "harness" in the context of English learning, I still don't quite get it. What kind of system would it be in the case of studying?

For example, a rule like "You can't move on to new words until you've reviewed today's 10 words" is a harness. Deciding to always reread your work yourself before moving on, instead of blindly accepting AI corrections, is the same idea. A checklist like "I won't accept corrections that don't meet these three criteria" or a sentence like "Please ask questions if you need more information" are also excellent, simple harnesses. Starting from here is perfectly sufficient.

So, you don't need to build an elaborate system; you can start creating one yourself today. I'll try adding the phrase "Ask a question first" to my prompts right away. It's reassuring to know that the AI won't just rush ahead on its own, but will stop and ask for clarification. The most frustrating thing is when it confidently answers without knowing the answer. We've got prompts, context, and harnesses, and they're all gradually coming together externally, but now that we've come this far, what's next?

🔁 Vol.5 Loop Engineering

We've finally reached the last layer: loop engineering. An engineer at a major AI company says his job is no longer to directly type prompts himself, but to write systems that continuously provide instructions. Apparently, before working in AI, he was engrossed in making fermented foods in Japan. Also, a renowned researcher reportedly completed a typing app for his daughter in about an hour with almost no input. This is a perfect example of a loop in motion.

The transition from making fermented foods to developing cutting-edge AI is so unexpected it made me chuckle. Waiting patiently for fermentation to occur and waiting for a loop to complete are perhaps quite similar after all. In both cases, it seems important to just watch without interfering. But what's truly amazing is that the app was completed in almost an hour with virtually no input. Does that mean the AI was checking and correcting everything on its own? How exactly does that work?

A loop is a mechanism that allows the "execute, verify, correct if necessary, and execute again" cycle to run without human intervention. With prompts, context, and harnesses in place, the AI can construct its own next instructions and continuously send them to itself. The human role shifts from issuing instructions each time to designing the system. However, this is not a complete handover. The essence lies in drawing the line between what to let go of and what to continue to control.

So, instead of leaving everything to AI, humans still have a say in the process. For example, with English learning, you could let AI handle the daily corrections and repetitive practice, but you'd still hold onto the goal—why you're learning English and what you ultimately want to achieve. If the AI decides the goal too, then it's not really my learning anymore. In that case, I think I could entrust it to it with peace of mind. By the way, is there only one type of loop like that?

🪆 Vol.6 Triple Nested Loop

Actually, there's a concept that involves three nested loops operating at different speeds. The innermost loop runs in minutes, with the AI rapidly repeating implementation, testing, and correction. The outermost loop runs from hours to days, with humans supervising, checking results, and fine-tuning the direction. And the outermost loop runs monthly, where feedback from friends and a select group of people is used to revise the entire strategy. One researcher has created a very clear diagram illustrating this.

It's like three overlapping loops of fast, normal, and slow speeds. It's kind of easy to understand if you imagine gears of different sizes meshing together. The small gears turn fast, and the large gears slowly determine the direction. The interesting thing is that only the outermost layer involves human reactions, not AI. I wonder how these three loops would work if I applied them to my own English learning?

The fastest loop is daily repetitive practice with the AI. You get your English composition corrected today and make corrections on the spot. The middle loop is once a week, when you review your correction log. The outermost loop is once a month, when you have a teacher look at your English or try using it in an online English conversation. Only the fastest loop can be left entirely to the AI. You must always have a human eye on the other two. This is what I meant by "holding on" earlier.

So, you divide your tasks into today, this week, and this month. I was also using AI-based proofreading daily, but I might have been skipping the weekly review and the monthly external feedback. I was probably satisfied with just doing it every day. Not relying solely on AI improves the learning process itself. But I wonder if this kind of thinking can be applied to other people's work, not just English learning or the world of engineering?

🌱 Vol.7 Summary — Loop Thinking Spreads to Non-Engineers

Absolutely. The four stages we've discussed so far are nested, with new layers being added on top of the old ones, rather than any of them becoming outdated. The inner layers continue to work even now. And this concept is spreading to administrative work as well. When a meeting ends, the AI automatically starts up, transcribes the meeting, identifies tasks, and even handles scheduling and sending out invitations for the necessary members. We're even seeing a system where people only need to click "attend."

The idea of people simply clicking a button to join seems like something out of the future, but it's already happening. It's amazing how a way of thinking born in the world of engineers can be directly applied to secretarial and administrative work. Organizing notes and scheduling after meetings really does take up a lot of time. This is surely relevant to our daily work and studies as well. But creating such a large system all at once seems difficult. Where should we even begin?

You don't need a big system right away. It's enough to start with a small, recurring loop, such as "having your work corrected by AI in the same format every day." For English learning, the initial form is a three-layered loop: today is repetitive practice with AI, this week is review, and this month is feedback from an external source. You can then gradually add prompts to indicate your purpose, context to reveal your weaknesses, and harnesses to encourage you to ask for clarification. There's no need to rush.

From prompts to loops, our interactions with AI are expanding outwards more and more. But the small loop we built today will surely serve as the foundation. Everything we've talked about in this series is connected to that. I think I'll start by asking for corrections on tomorrow's English composition in the usual way. If you'd rather read a detailed explanation than listen to a conversation, please also check out the original note article. I'll leave the link here.

👉 Read the original article
From Prompt to Loop (note Magazine)
https://note.com/gentle_hawk873/m/mc79175abd198