스크립트
00:00:00You cannot trust Claude to grade its own work, which is a huge problem, but one that this skill
00:00:05solves. It's called Claudex Loop, and the premise is simple. Instead of having Claude be responsible
00:00:09for planning, executing, and grading its own work, why don't we bring in Codex to also take a look
00:00:16at the plan and the execution of Claude and say, hey, this looks good, this doesn't, here's what I
00:00:22think you should change. Because one of the big problems with every single AI model out there
00:00:25is that they look at their own work very favorably. When I ask Claude how good the plan is that it
00:00:30created, it's going to say, this thing is awesome. So it's important we have systems in place where
00:00:35we can bring in a second pair of eyes to look at what the first model built and say, thumbs up,
00:00:40thumbs down, and why. So today, I'm not only going to give you the skill, I'm going to break down how
00:00:44it works under the hood, and we'll do a quick demo. Now, the skill is broken out into four phases,
00:00:48and the idea is you invoke it before you add some sort of feature, or if you're starting a brand new
00:00:53greenfield project. And the general idea is whatever the first model comes up with as a plan or whatever
00:00:58it executes on, we wait until the second model takes a look at it and says, hey, this looks great
00:01:02before we move forward on big things. Now, for the first phase, we're doing some reconnaissance. We're
00:01:07actually going to go out there on the web, Claude is going to scout out to see if the answers actually
00:01:10exist. We have the option to invoke deep research, which is the built-in dynamic workflow, if we want
00:01:15to get really deep with the answers we're receiving. After that, we go into the interrogation phase.
00:01:20Think of that as an enhanced plan mode set of questions, where we try to get on the same page
00:01:25with Claude before it creates its first plan. Secondly, this is where we bring in the second
00:01:28model. This is the review stage. So Claude code, after it comes up with a plan based on everything
00:01:34we've talked about and done our research on, it's going to create your standard plan.md.
00:01:38From there, Codex is going to review the plan in a read-only sandbox, and it's going to say,
00:01:43hey, either this is approved, or we need to revise x, y, and z. It's then going to send that answer
00:01:49with the revisions to Claude code. Claude code is going to take a look at it, say, I agree, I don't
00:01:55agree. And then it's going to send back its changes. This loop will continue up to five times. Now, I've
00:02:01never had it actually stall out at five times, or we didn't reach an approved state. However, I have
00:02:06it stopped at five, you can easily change that with the skill so that it doesn't get stuck in some weird
00:02:11endless loop, and you're just burning tokens forever and ever. This gives you like a very clear
00:02:15hard wall so you don't get stuck in that situation. Lastly, once Claude and Codex come to an agreement
00:02:20and a plan has been created, we then move into the build phase. Now, this skill gives you the ability
00:02:24to not just have Claude build, you can have Codex start building things as well. But whatever model
00:02:30builds it, the second model is going to take a look at what it actually created again before it moves
00:02:34forward with anything. And again, under the hood, there's a lot of different variables you can tune.
00:02:38Like I mentioned before, we have the review section, which is set to five. In the build section,
00:02:43when we have Codex take a look at what we've built, I've knocked that down to two loops again,
00:02:48just so we didn't get stuck in these sort of endless loop scenarios. And I haven't really run into
00:02:52issues where I felt like, Oh, I needed to bump this up. But you can play around with that. Further,
00:02:56you could play around with this more and actually like bring in a local model instead of Codex,
00:03:00if that's also how you want to approach this. Now, if you use the previous version of the skill,
00:03:04which was called Grill Me Codex, the changes you want to pay attention to with the Claudix loop
00:03:08are a more enhanced interrogation mode. So it's going deeper with its line of questioning. And in
00:03:13the execution phase, we've integrated Codex more. So again, we can get eyes on when it comes to the
00:03:18actual code that's been created. So next up is the actual demo. But before we jump into that,
00:03:23a quick word from today's sponsor, me. So I just released a completely updated version of my
00:03:28Claude code masterclass inside of Chase AI plus and it is the number one way to go from zero to AI dev,
00:03:33especially if you don't come from a technical background, we focus on real use cases. This is
00:03:38updated every single week. So if you're someone who wants to get better at this insane tool, and you
00:03:42just don't come from a software development background, this is for you. So if you want to
00:03:47check it out, there's a link to it in the pin comment. Now for today's demo, we're going to use
00:03:51Claudex loop to recreate Calendee. If you don't know what Calendee is, it's a scheduling web app,
00:03:56you give it a link to people, they can see your calendar, they can pick times, it automatically
00:04:00creates either a zoom link or a Google Meet. So it's something that a lot of people use and actually
00:04:05pay for. But I figured, hey, why don't we just create this ourselves and save us, you know, 10 bucks
00:04:09a month or whatever it is I'm paying, I should probably know that. So what I'm going to do is I just
00:04:13invoke forward slash Claudex loop, and I'm just going to give it a stream of consciousness and say
00:04:18something like, I want to use Claudex loop to essentially create our own version of Calendee.
00:04:25As of right now, I would probably just have it use Google Meet instead of zoom. But I'd want it tied
00:04:32with my calendar and essentially just recreate Calendee all the major features. So let's go ahead
00:04:38and do that. So at the beginning, we're in phase zero, which is the research phase. So it's saying,
00:04:41hey, how do you want to actually do this research? Either we do the web search, which is sort of your
00:04:46standard clause, going to send out a few sub agents, or do we want to go full blown deep research? Now
00:04:50with this skill, I have it pinned for Opus. So as you know, with deep research, if you do it on Fable
00:04:56and you just run slash deep research, it's going to call Fable sub agents, which can kind of go nuts
00:05:01on your usage. So as of now, I have it just do Opus off the bat, just to sort of help you out. Again,
00:05:05you can totally do that. It's recommending web, but honestly, I'm going to have it do deep just to see
00:05:10what it comes back with. It's going to show me the proposed deep research prompt and the questions it's
00:05:14actually trying to get answers to. So it needs to know about Google Calendar, Meet, scheduling domain
00:05:19pitfalls, and generally the stack. So if I prove that, we'll just say, go ahead and launch. And if I
00:05:25wanted to edit it, it's very simple to do. So Claude finished doing deep research and then created an
00:05:29assumptions ledger, which is essentially a list of everything it assumes you want to happen for this
00:05:36project. Now, after this, it's going to have more questions for us where we can kind of branch down different
00:05:40paths. But these are the things where it's like, Hey, I don't think there's a lot of question here.
00:05:44Now you can either just give this a thumbs up and it's going to do all this. And this is kind of get
00:05:48baked into the plan. Or you could say, Hey, actually, I want to change the stack or, Hey, I want to change
00:05:53how we're doing things with reminders and things of that nature. But for now, we're going to say, Hey,
00:05:57this is confirmed. And then we're going to move into sort of the, what it calls the load bearing
00:06:01tier. You know, we all love the term load bearing. That's now been appended to everything.
00:06:05Claude does. So now we're moving into those load bearing questions. The first question is,
00:06:09which Google account does your real calendar live on? For all of these questions, no matter your
00:06:13project, it's going to give you a series of questions alongside a recommendation. So if you have no clue,
00:06:17you can just go with recommended. But frankly, as a best practice, if you have no idea what Claude is
00:06:23telling you when it comes to like potential answers or like what is even going on, I highly suggest you
00:06:27go into this other section and basically just say something like explain this further. This is how
00:06:32you're actually going to get good when it comes to like AI and building things and not just being an
00:06:35accept monkey and hitting recommended, recommended over and over again. If you don't know, ask Claude
00:06:40to continue explaining it to you until you understand it. That's the only way you're actually going to
00:06:44learn, but kind of outside the scope of this video. So we're going to go with personal Gmail, and I'm
00:06:49actually going to skip ahead to after I've answered all these questions, because I think you kind of get
00:06:53how this phase works. Now, after we answer the load bearing questions, then we go into some
00:06:58cosmetic decisions that again, aren't really changing the base functionality of the application.
00:07:02In this scenario, it just lists them all out similar to the assumption ledger. So you can either say
00:07:08this is acceptable and it's going to just move forward with it, or you can say, hey, change one,
00:07:12change two, change whatever. I set it up this way so you can just sort of speed this process along.
00:07:16Now, after we answer those cosmetic questions, we then move into phase two. Claude is going to write
00:07:20that plan.markdown file, and then it's going to get sent to Codex using GPT 5.6 Sol. From there,
00:07:27they will have their back and forth for a maximum of five rounds, or until they reach an accepted
00:07:32verdict. So Claude and Codex went back and forth for five rounds. At the beginning, there were 27 issues
00:07:37that were brought up in round one, and we've gotten to round five, and the issue is there's still a few
00:07:43more problems. They haven't come to an accepted verdict. I mentioned earlier how I'd never had this
00:07:47happen before, so I'm kind of glad it did here. So we've hit five rounds. They haven't figured it out
00:07:51yet. There's still a few sort of just minor issues. What do you want to do? Well, you have options.
00:07:56You can either stop here, keep it as is. You can just accept the deadlock state, or you can extend
00:08:01it two rounds. So that's what we're going to do, and then we'll see what happens when they finally
00:08:05get on the same page. But nice to know and for you to see that even though we have sort of these hard
00:08:11stops, for example, five rounds, you aren't stuck with that. If you get to that point like we did here,
00:08:15you can very easily extend it. So after seven rounds, they've come to an agreement, and now it's going to
00:08:19ask you, who do you want to actually build this thing? So we can either have Claude build and codex
00:08:24takes a look, or we can flip it and codex builds and Claude takes a look. In certain situations,
00:08:29depending on what we're doing, it will also give you the option to do sort of a tandem build. So let's
00:08:33say we were creating something that required like asset generation or something like GPT images to
00:08:38brought into the equation. Well, it will also say like, hey, let's bring in codex for this specific
00:08:43portion of the project. But in this case, we're going to do Claude builds. And now it's going to
00:08:47start creating what it is called open book. So Claude could finish creating the calendar demo,
00:08:51which is what we're looking at here. So let's see if it actually works. And then we'll go back and
00:08:55actually dive into what codex added to this entire process. So we have the intro call and then a working
00:09:01session. So if we go to the intro call, we can then see a bunch of different times and schedules
00:09:07that are actually synced up to my Gmail calendar. So let's say I'm picking something for the 31st.
00:09:12And I picked 10am. You can just add my name, we'll put my email. And then we'll do confirm booking.
00:09:23And we can see the email got sent to us with a join link. And I can also see all my calendar,
00:09:28I also had it create just a simple artifact, kind of breaking down visually what codex added to this
00:09:33entire process. We talked about this earlier, that back and forth phase where Claude had the plan and
00:09:38codex made its, you know, adjustments or proposed adjustments that went all the way to
00:09:42seven rounds. And it started with 27 issues. And you can see each round, it got lower and lower until
00:09:48round seven, we finally had an approved verdict. And now here are some of the things that codex actually
00:09:52brought to Claude's attention during the planning phase, stuff like the double booking constraint
00:09:57could not compile things like OAuth connect flow having issues. There's also problems with concurrency
00:10:03that two reschedules the same booking could both succeed and on and on and on a lot of edge case
00:10:08related stuff. Now in the build phase, Claude implemented that improved plan we had after the
00:10:12seven rounds. And then we had a new codex session pop up with completely fresh memory, new context
00:10:17window, it hadn't read the plan. And then it read the code against the actual spec, what was created
00:10:23versus what was actually planned. So it came back with 23 findings, 19 were accepted and fixed, four were
00:10:29rejected. So here's some of the things that codex found in this build phase, the time grid drifted after
00:10:34every meeting, the management token was sitting in plain text, all the events blocked the wrong hours.
00:10:39And again, more and more things that, again, you could argue our edge cases, but it would have
00:10:44taken Claude a long time to find them. So if codex had never been in the room from the beginning,
00:10:48what would this have looked like? Well, we would have had broken features, we would have bookings that
00:10:53exist in the database and nowhere else and so on. Now, in reality, would this have been the case if we
00:10:58just relied on Claude? Probably the beginning, we just would have taken some further iterations and we
00:11:03eventually would get to something that probably worked. But with codex, we found a lot of these issues in the
00:11:08planning phase. So we didn't have to burn tokens and then burn more tokens after the fact. And we were
00:11:14able to test these things with codex, take a look at these with codex before moving to production. So all in all,
00:11:21this saves you a lot of time and money. So that's the Claudex loop in action. If you want to get your hands on
00:11:26this, I will put a link to it in the pinned comment. And besides that, I'll see you around.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기