AIに 指示する 時代の終わり|ループエンジニアリング徹底解説 — Transcript
Full transcript
- 0:00Mr. Ta, I've been hearing the term "
- 0:01harness engineering" a lot lately, and
- 0:03our engineers have been working hard on
- 0:05putting together harnesses recently,
- 0:07right? That's
- 0:08right. But I thought it was already
- 0:10outdated.
- 0:10Huh, harness engineering? We're
- 0:12entering the next phase. Actually,
- 0:14really? I thought it had just come out
- 0:16recently. There's a type of
- 0:17engineering that's becoming popular
- 0:19right now.
- 0:19Loop engineering. And there's another
- 0:21new one emerging.
- 0:22Yes. It's been a hot topic on X
- 0:24recently, and it's also starting to be
- 0:26mentioned in places like Anthropik.
- 0:28What kind of thing is it? It's
- 0:29something called prompter engineering
- 0:31itself. You're not doing it. You're not
- 0:33even
- 0:33giving instructions to the AI?
- 0:34Yes. You're not giving instructions.
- 0:36How are you doing it? You're having the
- 0:38AI give instructions.
- 0:40Wait a minute. I can't keep up. The
- 0:42concepts are getting more difficult, so
- 0:44I, a travel engineer, would like to
- 0:46explain it in a way that even beginners
- 0:48can understand. Is it
- 0:49another development method using AI?
- 0:52Yes. I guess it's a method for
- 0:54AI-driven development. Group
- 0:55engineering. It suddenly started to
- 0:58catch on, but this was in June 2026.
- 1:00It's been a while since then.
- 1:03Everyone started saying exactly the
- 1:05same thing. The first person is Peter
- 1:07Steinberger. He's the developer of
- 1:08OpenClaw, and he's apparently now at
- 1:10OpenAir. His idea that we shouldn't
- 1:12give prompts to agents anymore, but
- 1:14should design loops that give prompts
- 1:15to agents, went viral on Flex. The
- 1:18second person is Boris Kearney. He's
- 1:20the person in charge of the
- 1:22Anthropocene Clod Code. Yeah.
- 1:24This person said that his job is to
- 1:26write loops. He deleted his IDE six
- 1:28months ago.
- 1:29Yes. What's an IDE? It's a
- 1:32development environment. It's an eliter
- 1:34or integrated environment for
- 1:36development, and you usually enter
- 1:38prompts and give instructions there,
- 1:40but he's not doing that. The third
- 1:43person is D. Oma. He's the director of
- 1:46engineering at Google. He named it "
- 1:48P-Engineering" on his blog and
- 1:50systematized it on his blog. Around
- 1:52this time, things started to catch on
- 1:54and everyone started talking about loop
- 1:56engineering. Yeah.
- 1:57People at the forefront of AI are
- 1:59saying that if you don't give
- 2:00instructions to AI, you shouldn't be
- 2:02typing "prompt". That's
- 2:03right. It's not about typing prompts,
- 2:06it's about designing loops. That's
- 2:08where everyone started to move. The era
- 2:10of humans giving AI sequential
- 2:12instructions is over. From now on, it's
- 2:14about designing the mechanism itself
- 2:16for giving instructions. I actually did
- 2:18group engineering. I tried it myself.
- 2:21I'll explain that a little later. It's
- 2:22a development style. It's not like this
- 2:25concept of loop engineering suddenly
- 2:27emerged. It's more like a trend of
- 2:30engineering that has been powered up by
- 2:32prompt engineering. Prompt engineering
- 2:36was the first thing that was said.
- 2:38There was also talk about how to use
- 2:40blomp skills. Yes. I think this was
- 2:42probably the first time it was said,
- 2:44and it was said until around 2024, and
- 2:46the design concept was to increase the
- 2:48quality of each proton. And the next
- 2:51one came out in 2025. It's context
- 2:53engineering. This contest engineering
- 2:55is not just about giving instructions
- 2:57as an information environment to show
- 2:59to the model, but also about everything
- 3:01including the information behind those
- 3:03instructions. Harness engineering came
- 3:05out in 2026. Isn't it just new? It
- 3:09came out this year, but it's the next
- 3:11stage. Harness engineering is about
- 3:15designing rules, constraints, and
- 3:17verifications to see if AI can run
- 3:19perfectly. I've been working on harness
- 3:22engineering recently, but the new one
- 3:24is loop engineering. This means
- 3:25creating a mechanism to implement using
- 3:28harnesses from June 2026. Yeah.
- 3:30Development methods have changed so
- 3:32much since the emergence of context
- 3:33engineering. I can't keep up anymore. I
- 3:35think chat was the basis when we were
- 3:37doing front-end engineering and context
- 3:39engineering. When we entered context
- 3:41engineering, AI agents became the basis
- 3:43. And that's where AI agent-based loop
- 3:46engineering came in. I don't want you
- 3:49to misunderstand that it's not that
- 3:51we've switched to a different concept,
- 3:53but rather that we've gradually
- 3:54incorporated the existing concepts.
- 3:56Based on that, it's like wrapping it up
- 3:59one more time. Fronts are important,
- 4:01contests are important, and harnesses
- 4:04are also necessary. Engineering now.
- 4:06That's
- 4:06right. Originally, context engineering
- 4:08was based on prompts, and the question
- 4:10was what kind of information to add to
- 4:12the prompts. Harness engineering was
- 4:14the idea of thinking about how to make
- 4:16it run autonomously based on that added
- 4:18information. Loop engineering is the
- 4:20idea of running it autonomously, not by
- 4:23running a single task autonomously, but
- 4:25by looping it repeatedly to complete a
- 4:27larger task. Isn't that going too far?
- 4:31It starts
- 4:32with the definition of loop engineering
- 4:34. Loop engineering is not about giving
- 4:36instructions, but about creating a
- 4:38mechanism to give instructions. This is
- 4:40the basis. It's the idea of designing a
- 4:43loop itself that continues to give
- 4:45prompts to the AI coding agent, instead
- 4:47of manually typing prompts to it. In
- 4:49the past, humans would type prompts.
- 4:52The AI would respond, and humans would
- 4:54give the next instructions. The AI
- 4:55would then work again, and this would
- 4:57be repeated. That's the past approach.
- 5:00Loop engineering involves humans
- 5:01defining the goal. Once the goal is
- 5:04defined, A automatically gives
- 5:05instructions within that system,
- 5:07verifies the results themselves,
- 5:09decides on the next task, and continues
- 5:11until the goal is achieved. Does this
- 5:14mean that
- 5:14humans create a system where AI gives
- 5:16instructions to AI?
- 5:17Yes. I
- 5:19think it would be easier to understand
- 5:21if we were talking about a team of
- 5:22people. Up until now, I was acting as a
- 5:25player, like a playing manager. I would
- 5:27give instructions to my juniors, verify
- 5:29their results, and then say, "Okay,
- 5:31let's move on to the next step." The
- 5:33entire system for loop engineering
- 5:35development is defined here, so please
- 5:37create functions within this framework.
- 5:40The team thinks and moves forward, and
- 5:42that's how it works. Does it
- 5:43feel like the management burden has
- 5:45increased?
- 5:46Yes. It's gotten a lot more involved in
- 5:48the management field. It's designing
- 5:50how to get the team moving. In cooking,
- 5:53there are chefs, stonemasons, and other
- 5:55people, but up until now, I was the one
- 5:58watching over them and having them help
- 6:00me out in the room, but in group
- 6:02engineering, the owner would say, "
- 6:04Please make this kind of menu," and
- 6:06design all the rules for the kitchen.
- 6:08The idea is that a chef or something
- 6:10like that has become AI and starts
- 6:12moving on its own. It's a way of
- 6:13thinking about designing a team. When
- 6:15people talk about loop engineering,
- 6:17they often think of regular guidance
- 6:19and retraining. It's easy to mistake it
- 6:22for a loop, but loop engineering
- 6:23doesn't just have the model repeat the
- 6:25same process; it looks at the state and
- 6:28decides on the next action. The idea of
- 6:31loop engineering is to make decisions
- 6:34while looping forward toward the goal,
- 6:36rather than doing routine work. Is that
- 6:39even possible? It's like giving AI a
- 6:42personality, creating a team, and
- 6:44creating a state where it can properly
- 6:46reduce in a loop. It even sounds scary.
- 6:48In fact, it's not just a buzzword that
- 6:51came to mind, but has a solid research
- 6:53foundation. There are two: the idea of
- 6:56react and reflection. React is a
- 6:58pattern of thinking, acting, observing,
- 7:00and thinking again, repeated over and
- 7:02over. In programming terms, it's like
- 7:04writing, executing, checking for errors
- 7:06, fixing them, and passing the test.
- 7:08This idea of repeating this process is
- 7:10based on the idea of react. Reflection
- 7:12is a pattern of verbalizing and
- 7:14recording failures and using them for
- 7:16the next try. Based on these, we have
- 7:19loop ending. It's a system that builds
- 7:21a mechanism for learning from mistakes
- 7:23that you can fix on your own. It's like
- 7:26systematizing what people do. We've
- 7:28talked about the concept, but from here
- 7:30on, we'll get a little closer to the
- 7:32implementation. There are two layers to
- 7:35a loop. I think understanding this will
- 7:38help you understand a lot better. The
- 7:41two are the inner loop and the outer
- 7:42loop. It's a system where you verify
- 7:46things yourself within a single task.
- 7:48Is
- 7:49this the inner loop?
- 7:50Yes. I think many of you often have
- 7:54complaints about AI, like when you ask
- 7:56it to fix something, the AI says it's
- 7:58done. But when you look at it, it's not
- 8:02fixed or it's still broken. You can
- 8:04verify that by running a loop. You
- 8:07don't just say it did it, you have it
- 8:08check to see if it's fixed?
- 8:10Yes. You edit, create a test, run the
- 8:12test, detect failures, detect other
- 8:15patterns, fix them, re-run it, and
- 8:17check if it passes. This loop is
- 8:19repeated. The
- 8:21inner loop is the one that verifies and
- 8:23fixes things until it's complete. In
- 8:25the conventional format, people do it,
- 8:28people give instructions, AI does it,
- 8:30and people check it. Instead, people
- 8:32give instructions, AI executes it, AI
- 8:34checks it, and if it doesn't work, it
- 8:36goes back to the previous AI, and
- 8:38that's repeated all the way to the pole
- 8:40and finally goes back. This is what's
- 8:42called the inner loop. The key point
- 8:45here is to do verification in detail.
- 8:48For example, it's better to clarify
- 8:50what you want to verify. Even if you
- 8:55create something that only performs
- 8:57testing, there are other things like
- 8:59security risks, code errors, and slight
- 9:01inconsistencies in the specifications,
- 9:03so if you review each of them, the
- 9:05verification will be more effective.
- 9:09Are there many AIs that perform testing
- 9:11and AIs that check security? Yes. By
- 9:14doing that, it's like you can't proceed
- 9:16to the next gate after four gates. Yes.
- 9:19It's an AI team. Yes. Next is the outer
- 9:22loop. The current inner loop is a small
- 9:24single task. Since this single task
- 9:26alone cannot meet the requirements, you
- 9:28line up the functions and run them all
- 9:31over and over again. It's like running
- 9:33a session. If it takes one session to
- 9:36create one function, and you continue
- 9:38to do that over time, and gradually
- 9:40move on to the next task, you're
- 9:42creating a loop, right? Yes, that's
- 9:45right. The point is to learn more by
- 9:47repeating. If there was a mistake in
- 9:51the first task, you can design it so
- 9:53that you can apply the lesson to the
- 9:55next task,
- 9:56and then you'll be able to use that
- 9:58lesson to talk about different tasks.
- 10:00It's like properly
- 10:02passing on the PDC. It's like designing
- 10:04to avoid repeating the same problem.
- 10:06Yes. It's like developing people. In
- 10:09Loop Engineering, written by Ali
- 10:11Diomani, there's a concept called the 5
- 10:13+ 1 components that make up a loop. He
- 10:16organized it so that a properly
- 10:17functioning loop consists of five
- 10:19components plus memory. I think you
- 10:21might not understand what I'm talking
- 10:23about, so let me explain. The first is
- 10:26automation. The role of automation is
- 10:29to periodically discover tasks and
- 10:31start agents. This is the core of the
- 10:34loop. The second is the work tree. It's
- 10:37the idea of properly separating work
- 10:40areas so that multiple agents don't
- 10:42collide. So that two AIs don't do the
- 10:44same thing, right? Well, if they work
- 10:46in the same place, the environments
- 10:47will clash, so let's create an
- 10:49environment where they can work in
- 10:50parallel. The third is skills. This
- 10:54project will be permanently persisted,
- 10:56which might be the idea behind
- 10:58harnesses. The fourth is flag
- 11:00connectors. Things like Habooslack DV.
- 11:04Let's prepare things to connect to
- 11:06dependent tools and properly connect to
- 11:08the outside world. The first is
- 11:11subagents. As I mentioned earlier,
- 11:14these are necessary for the division of
- 11:16labor in a team, so let's separate the
- 11:18roles of creation and verification. And
- 11:20then there's the plus 1, memory. The
- 11:23key point is that the state is
- 11:24persisted outside the conversation, not
- 11:26within the session. If we process
- 11:31something within the session, the
- 11:33processing ends there, so the idea is
- 11:35to have it persisted externally. For
- 11:38example, let's record any missing
- 11:40failures outside the task, not within
- 11:42it. That's
- 11:43right. This is the 5 + 1 approach.
- 11:45Among these, I think the fifth point,
- 11:47the subagent, is something that needs
- 11:49to be designed carefully. It's
- 11:52closer to the idea of a harness than a
- 11:54subagent loop, but in order to ensure
- 11:56the quality of what AI creates, it's
- 11:58very important to separate the roles of
- 12:00the agents. I think I talked about the
- 12:03harness in the anthropics before, but
- 12:05it's important to clearly separate the
- 12:07roles of creator and checker. In terms
- 12:09of what I've talked about on this
- 12:11channel, first there's the planner. The
- 12:13planner makes the plan, then there's
- 12:15the generator. The implementer writes
- 12:17the code, and the elevator, the checker
- 12:19, checks and provides feedback, and
- 12:21this cycle goes on and on. If you work
- 12:23on this part carefully, the quality
- 12:25tends to improve. In a previous paper,
- 12:27there were results showing that
- 12:29dividing the roles rather than having a
- 12:30single agent do it leads to better
- 12:32results. That's
- 12:33right. Designing the roles of the team
- 12:35is key. Another thing is to have a
- 12:37state, but since it's passed on to the
- 12:40subagents one after another, even if a
- 12:42subagent makes a mistake, it won't be
- 12:44passed on to the next one. That means
- 12:46that by writing out the failed tasks in
- 12:48the subagent to an external file, you
- 12:50can write out the status to the file to
- 12:52indicate whether it failed or succeeded
- 12:54and move on to the next task. If you
- 12:56make sure to write out completed tasks,
- 12:58tasks in progress, and things learned,
- 13:00the quality will gradually improve. So
- 13:03that's how far you've
- 13:04come in your work. It's like you're
- 13:06designing an organization that lets you
- 13:08learn and not let the next one go.
- 13:09That's
- 13:10right. I'll
- 13:11tell you something even more important.
- 13:14Some people might think that the loop
- 13:17is a place where some amazing genius,
- 13:19like AI, acts on its own, but to put it
- 13:21simply, the loop, the team that I
- 13:23mentioned earlier, is not really run by
- 13:26super engineers. It's more like a group
- 13:29of juniors who work tenaciously. I see.
- 13:31If the instructions are vague, they'll
- 13:33get together, and if the goal is not
- 13:35clear, they might go in the wrong
- 13:36direction, and if you don't set any
- 13:37restrictions, they'll just keep trying
- 13:39endlessly. So it's like having a lot of
- 13:42juniors. That's right. It's good to
- 13:44think of it as a team of juniors. It's
- 13:46like defining the rules clearly. There
- 13:48are four important things to keep in
- 13:49mind when defining rules. The first is
- 13:51a clear task definition. Can you
- 13:53clearly state when it's complete in one
- 13:55minute? Make sure that everyone
- 13:57understands what your goal is. If the
- 13:59goal definition is vague, the outcome
- 14:01will also be vague. The second is a
- 14:03safe work environment. As I mentioned
- 14:05in Work 3 earlier, the design should be
- 14:07such that even if something goes wrong,
- 14:09the environment is isolated and the
- 14:11whole thing won't break. It's better to
- 14:13design this in advance. The third is
- 14:16whether or not you can see the success
- 14:17visually. This means using tests,
- 14:19linting, and type checking to
- 14:21automatically determine the overall
- 14:23cost. It's like having a solid cariness
- 14:25system in place. By doing this, it
- 14:26becomes easier to see whether or not
- 14:28something has clearly been successful.
- 14:30The fourth is a hard stop. This
- 14:32prevents infinite loops and runaways.
- 14:34If you keep working until this quality
- 14:36is met, you'll just keep working
- 14:38endlessly. You won't have any idea what
- 14:40the outcome will be. You need to set it
- 14:42up so that if you get stuck, you can
- 14:44ask your superiors. It's something like
- 14:47including the maximum number of retries
- 14:49, how much time it should take, and if
- 14:51the cost exceeds a certain level, stop,
- 14:54or confirm. To prevent them from
- 14:56running too fast and collapsing,
- 14:58juniors tend to take on too much on
- 15:00their own and keep doing tasks
- 15:02indefinitely, right? I think there are
- 15:05rules like "if you're having trouble
- 15:07with time, ask someone" and so on. In
- 15:10the same way, it's important to clearly
- 15:12define what and how much you should get
- 15:14stuck on and what you should fail at in
- 15:16the conjunctive case. I thought it was
- 15:18about team and organizational
- 15:20management. Since we were talking about
- 15:23next year, I'll actually talk about the
- 15:25loop I created from here. In general
- 15:27terms, I created a skill called Epic
- 15:29Flow for the outer loop and a skill
- 15:31called Transfer Flow for the inner loop
- 15:33. I'll start by explaining the inner
- 15:35loop transfer flow. Since it transfers,
- 15:37the implementer implements it. There is
- 15:39also a tester as another subagent. The
- 15:41tester tests. If it passes, the next
- 15:43step is a multi-axis review. The
- 15:47reviewers I created are a usage checker
- 15:49, a code checker, and a security
- 15:51checker. There are three reviewers.
- 15:54Once it passes the four-axis review, it
- 15:56is returned to the implementer, who
- 15:59then submits a full request. Let me
- 16:01explain in more detail. The one-cycle
- 16:04flow is called the epic flow from the
- 16:06middle loop. An epic is a large feature
- 16:09that can be given a name. When you
- 16:12start the pick flow and enter the
- 16:14requirements there, the planner is
- 16:15started first. The planner breaks down
- 16:18the major requirements into
- 16:19sub-requirements. In creating the
- 16:21sub-requirements, dependencies,
- 16:23acceptance conditions, and objectives
- 16:25for each sub-requirement are determined
- 16:27. There are several things to do to
- 16:29create a large feature, so the runner
- 16:31breaks them down and determines them.
- 16:33Is this okay for now? Here, we will
- 16:35send in a merchant. When the human
- 16:37merchant
- 16:38sends the merchant, the main agent will
- 16:40give instructions one by one for each
- 16:42cycle. The one-cycle flow we saw
- 16:43earlier is used when giving
- 16:45instructions. In the one-cycle flow,
- 16:46one task is completed one after another
- 16:48.
- 16:48Yes. The one-cycle flow is created, and
- 16:50when it comes to completing tasks, the
- 16:52main agent gives instructions to
- 16:53complete them in the one-cycle flow.
- 16:55The
- 16:55main agent passes one task to the
- 16:57developer and implementer. After one
- 17:00week, we come back. Then, we make it an
- 17:03epic branch. By repeating this process,
- 17:07this entire function is completed.
- 17:10Finally, this entire epic pull request
- 17:12is completed, so a human checks it one
- 17:14more time, and then the initial budget
- 17:16and completion checks are done. This is
- 17:19the only part that humans do. The AI
- 17:22checks whether each of the smaller
- 17:23decomposed parts is done. If it's OK,
- 17:27it moves on to the next loop. The
- 17:29finished product is then checked and
- 17:31approved. That's right. The check is
- 17:34done on the assumption that the quality
- 17:35of the work is guaranteed. Is there a
- 17:37chance that the finished
- 17:38product will be far from what was
- 17:40expected? The
- 17:41accuracy is incredibly high.
- 17:42Yes, that's right.
- 17:43Yes. First, we hand over the major
- 17:45requirements, but if there are any
- 17:47ambiguous points, they ask questions.
- 17:49The planner asks questions, and this is
- 17:51what happens. Because we refine the
- 17:53requirements. Once they understand, the
- 17:55planner starts working on the task over
- 17:57. Because that's done well, a proper
- 18:01product is created.
- 18:02Yes. Yes, that's right. Also, under
- 18:04this main part, we have a harness
- 18:06properly drawn, so the quality check
- 18:08method and design are all handed over
- 18:11in advance in the harness, so it's even
- 18:13harder to get lost.
- 18:14How much time does it take? In the
- 18:16current phase 123, the plan can be
- 18:18completed in 5 to 10 minutes, but from
- 18:20there, the development confirmation
- 18:22takes quite a while, depending on the
- 18:24size of the previous day.
- 18:25Yeah. Yeah. Yeah.
- 18:27And we often hit the limit. As
- 18:28Hikomo-san said, "I hit the limit again
- 18:31." Recently, I've been making an epic
- 18:33before going to bed, executing it, and
- 18:35waking up in the morning, and it's
- 18:37finished. A great fairy gave me a shot.
- 18:39As I mentioned earlier about the
- 18:41planner, there are 6 sub-agents. They
- 18:43are planning, implementation testing,
- 18:45usage confirmation, and security
- 18:46confirmation. I'll tell you what kind
- 18:48of design each one has. Earlier, we
- 18:50were talking about using Q a lot, right
- 18:52? We use different models. The planner
- 18:55uses Opus. This involves breaking down
- 18:58tasks from the epic, determining
- 18:59dependencies, acceptance criteria, and
- 19:01areas of concern, as well as preparing
- 19:03objectives and security checks. There
- 19:05are some areas where security is not
- 19:07necessary, so we design here to
- 19:08determine whether it is necessary or
- 19:10not. Next is the implementation team.
- 19:12We use Sonnet for this. The
- 19:14implementation team has put together a
- 19:16solid harness, and we ask them to
- 19:17proceed with the description of the DD
- 19:19bull request based on the harness. The
- 19:21tester. This is also one of them. They
- 19:23check the implementation based on the
- 19:25implementation criteria. This is not
- 19:27just a single debug, but we also have
- 19:29them check the screen using a
- 19:30playwright. They then check whether it
- 19:33is okay, so to check whether the
- 19:35operation is working properly, it is
- 19:37not just the code but also the
- 19:38operation. We use a playwright to
- 19:42launch the screen in a browser and
- 19:44check it. The usage checker. We set the
- 19:48implementation criteria, but they check
- 19:50whether the implementation criteria are
- 19:52sufficient for the purpose and intent,
- 19:54and whether the scope of impact has
- 19:56spread to other areas. If there is a
- 19:58gap, we say that it is missing and
- 20:00return it. The last is the code
- 20:02reviewer. This is also done using
- 20:04Sonnet. If there are no issues, we'll
- 20:06just leave it as is. If there are any
- 20:08issues, we'll point them out right away
- 20:10. It's the same as human code reviews.
- 20:15Security checks. This is important, so
- 20:18we've used Opus. We're keeping a close
- 20:20eye on overall security, but we try not
- 20:22to start it up where it's not needed,
- 20:24and it only starts up when we're
- 20:26writing code. So, each of the six
- 20:29subagents starts up. This is how we set
- 20:32up an independent response. If I start
- 20:35talking about the details, well, there
- 20:37are things I want to say, but I
- 20:38gradually lose track of what I'm saying
- 20:40, and some people might think, "We
- 20:42don't need engineers anymore." But
- 20:44actually, when I ask juniors, they ask
- 20:46me how I can put together so many
- 20:48people. As I've been saying, it's
- 20:49because I have experience running teams
- 20:51of people. I've been working on a lot
- 20:54of development-style flows, such as how
- 20:56to make teams work when working on
- 20:58projects. So it just came naturally to
- 21:01me.
- 21:01Yeah. That's what AI has replaced. Well
- 21:04, when you think about the skills
- 21:05required for a loop engineer to work
- 21:07together, the first thing is how to
- 21:09create software as a GM. There are also
- 21:11a lot of things to consider, so if you
- 21:12learn these things, it will be easier
- 21:14to understand. Then there are project
- 21:17management and testing techniques. You
- 21:21also need to think about how to ensure
- 21:23quality. If you have upstream skills,
- 21:25it will be easier to work together. On
- 21:28the other hand, you are someone who has
- 21:30been running a team in software
- 21:31development. I think it would be
- 21:34difficult to know what kind of roles to
- 21:36create and what kind of loop to use. I
- 21:39think there are many things that need
- 21:41to be added, such as whether to perform
- 21:43verification or define a design
- 21:45document. However, it is not enough to
- 21:47think that this is the only correct
- 21:48loop. In software development, the
- 21:51method changes slightly depending on
- 21:53the site and requirements, and it needs
- 21:55to be flexible, so the way to work
- 21:57together needs to be considered
- 21:59according to each project. If you are
- 22:02going to work together all the time,
- 22:04you need to create a situation where
- 22:06people can run it. That is why studying
- 22:08is important.
- 22:09If anyone watching this is interested,
- 22:11can I consult with you?
- 22:14In the future, I'm thinking of reviving
- 22:16live streaming and such, so please feel
- 22:18free to ask me there as well. Also, for
- 22:21those who watched this video, I've
- 22:23created a small e-book called Loop
- 22:25Engineering. Yes. That's right. I'll
- 22:29hand it out, so if you register on LINE
- 22:31, I'll send it to you as soon as
- 22:32possible, so if you want to know more,
- 22:34please register on LINE. Yes. So you've
- 22:36even created an e-book? Yes. I've
- 22:39recently become a bit addicted to
- 22:40creating e-books. It's like a
- 22:42new hobby, so if you're interested,
- 22:44please register on LINE and check out
- 22:46the book. Yes. Mr. Hon, am I cut out to
- 22:49be an engineer too?
- 22:49If you're not sure, I'll take the test.
- 22:51I'd like to
- 22:52try it. You can easily take the test
- 22:53from LINE. Everyone, please try the
- 22:55engineer aptitude test from LINE in the
- 22:57description.
About this transcript
This page contains the full transcript of AIに 指示する 時代の終わり|ループエンジニアリング徹底解説 by エンジニア転職チャンネル【RUNTEQ公式】, generated from the public captions YouTube serves with the video. The transcript has 4,159 words across 660 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.