I Updated /grill-me And Solved Claude Code
CChase AI
Computing/SoftwareSmall Business/StartupsInternet Technology
Transcript
00:00:00Plan mode is not enough. Skills like Matt Pocock's Grill Me or even larger orchestration layers like
00:00:06GSD or superpowers are all trying to solve the same problem. Take that fuzzy idea in your head
00:00:11and turn it into something Claude Code can actually build. But no matter what path you go
00:00:16down or what skill you choose, they all run into the exact same problem. You're relying on a single
00:00:21model to not only plan and build, you're relying on a single model to actually grade its own work.
00:00:26So when you ask Claude, hey, was this the optimal path forward? What's it going to say? Well,
00:00:31it's going to say it was great no matter what you did. And this is a problem because if you do not
00:00:35come from a technical background, you don't actually know if what Claude wrote actually makes sense.
00:00:41But in this video, I'm going to show you how to fix that. We're going to build upon Matt Pocock's
00:00:45Grill Me skill and we are going to bolt on an adversarial code review from Codex. But it's a
00:00:51code review that goes well beyond the Codex plugin you've seen in the past. This code review is
00:00:55iterative. Claude Code and Codex are going to be talking to one another through multiple rounds
00:01:00to get you to a place where both leading AI tools sign off on your plan. So you actually can feel
00:01:07confident that what Claude Code came up with actually makes sense. And with this skill, you're going to be
00:01:12able to start every project with two things. One, a plan that you actually understand. And two,
00:01:18a plan that multiple AI tools have signed off on. So what you're going to get today are two skills for me.
00:01:23And both of those skills are built on the backs of what Matt Pocock gives us here in his GitHub repo.
00:01:28He has two skills, Grill Me and Grill with Docs. The two skills I'm going to give you are Grill Me Codex
00:01:35and Grill with Docs Codex. So what's happening? Well, Grill Me and Grill with Docs are essentially a plan
00:01:41mode on steroids. Just like GSD, just like superpowers, it takes it a step further. The questions it asks are
00:01:48deeper. It's going to give you better insight into what you're actually trying to build because whether you
00:01:53want to admit it or not, you probably suck at actually articulating what you want. And if you
00:01:57can't articulate what you want to Claude Code in the beginning, you're going to have a lot of
00:02:01assumptions on the AI side, which give you a mediocre product on the back end. So Grill Me and Grill with
00:02:07Docs give you better outputs by going deeper in the planning phase to make sure you're all on the same
00:02:12page. What my skills are going to give you is a second phase to that, where after you and Claude
00:02:19Codex have gotten on the same page, Codex comes in and says, hey, that makes sense. That doesn't fix
00:02:24this, fix that. And then Claude Code and Codex go back and forth. And I think this is important because
00:02:28stuff like Grill Me, GSD and superpowers, they identified this gap right here, this gap between you
00:02:34and Claude Code, where you have an idea, you can articulate it, we're going to go back and
00:02:38forth, we're going to get on the same page, right? Grill Me, you know, GM is perfect for this.
00:02:44The problem is, even if you and Claude Codex are on the same page, does that mean we are automatically
00:02:51on a journey to this place of optimal code, where this is what should actually be built? Maybe, maybe
00:02:57not, who's to say? You probably aren't to say, are you an expert software engineer? You might be,
00:03:03but I would guess most of the audience watching this does not fall into that camp.
00:03:08And stuff like Matt Pocock's thing, as great as it is, I mean, like, it's skills for real engineers.
00:03:13Are you a real engineer? Probably not. Maybe you are. If you're not, will you fall into the problem
00:03:19where you can't even evaluate what Claude Codex has written? Even if you're on the same page,
00:03:23it could be trash, it could be amazing, who knows? And the other issue is, you can't judge it,
00:03:28and neither can Claude Codex because Claude Codex, and this is something Anthropic themselves have said,
00:03:34is very nice and talks very well about the code it has written, right? You ask Claude Codex to judge
00:03:40what it's written, it's like, oh yeah, sick, A+. So, are they like a reliable narrator and a reliable
00:03:46evaluator in this case? No, they're not. So, if you don't know what's going on, and we can't
00:03:50necessarily trust Claude Codex, where does that leave us? Well, we have this gap here then, right?
00:03:56We have this gap between Claude Code and quote-unquote optimal code. And so, the obvious solution is,
00:04:02well, let's bring in a third party, a neutral third party to take a look at our plan. In comes Codex.
00:04:09And this Codex review is what I've added to Pocock skills, and it's what I'm going to be giving you today.
00:04:16So, the first half is exactly the same as GrillMe. Questions back and forth, we get this plan going
00:04:21together, everything is nice and neat right here. And once we have the plan all set in stone, well,
00:04:27then Codex is going to come in, it's going to see what Claude Codex has come up with and say,
00:04:32this looks good, this looks bad, what do you think? Claude Code is going to take a look at it and say,
00:04:36oh, that makes sense, let's fix that, here's what I did, take a look again, Codex. And it's going to
00:04:41go through a cycle of like, well, it maxes out at five turns, you can easily edit that, but it's going
00:04:48to have five back and forth, which is a little bit different than the standard adversarial review
00:04:52Codex plugin, because it's more iterative. And the idea being, if they go back and forth enough
00:04:57times, we'll eventually get to a place, hopefully sooner than five turns, where they're both like,
00:05:01hey, thumbs up, it's good to go, push forward. So all's that to say is what I'm giving you today
00:05:09is meant to fix this gap right here. This gap between Claude Code and the optimal code that you
00:05:16and I will struggle to identify because we are not expert software engineers and Claude Code can't
00:05:21be trusted to do it to a certain extent. So that's what we're covering. And now we're all on the same page.
00:05:28But before we hop into the demo, a quick word from today's sponsor, me. So as you know,
00:05:33Chase AI Plus is the home of my Claude Code masterclass. And it is the number one way to go
00:05:37from zero to AI dev, especially if you do not come from a technical background. We focus on real use
00:05:42cases. And I recently added the Claude OS masterclass there as well. So if you're like, hey, I also want to
00:05:49learn how to integrate things like Obsidian and create a full command center. This is the place for
00:05:54you. You can find a link to it in the pin comment. So for today's demo, we're going to add a new page
00:05:59to our website. So this is the website for my AI agency. And the new page is going to give people
00:06:05access to some exclusive skills. And to get access to this page, when they click on it, they're going
00:06:11to have to add their email. So it's somewhat gated, we grab their email, then they have access to the
00:06:16things they can download. Now the email needs to then get handled with our database, which already exists.
00:06:22So we're not just creating some feature from thin air, it needs to take a look at the code base that
00:06:27already exists and make it coherent. So this is the prompt I'm giving Claude code run grill me codex.
00:06:32I want to add a email capture gate to the site that unlocks the grill me codex Claude code skill.
00:06:38If visitor lands on a page where the skill download is blurred behind an overlay,
00:06:42they enter their email to unlock it and their email is stored. And then I gave it some additional context.
00:06:49So the first part is going to be the grill me skill. It's the exact same grill me part as
00:06:56Matt Pocock's one, the one we're kind of building off of. So that part is the same.
00:07:00And once we go through all the questions and codex will come in. So after I looked through
00:07:03the code base, it's now asking me the first question and saying, how real is this gate
00:07:07when it comes to the blur? Is it a cosmetic thing or is it actually going to be enforced?
00:07:11And just like with grill me, anytime it asks you a question and gives you some potential answers,
00:07:16it also gives its recommendation and why. So for this one, it's just going to be cosmetic.
00:07:21It's a free skill. The goal here is just to capture the email. So we're just going to say,
00:07:25cosmetic is fine. File is free anyway. Next is asking about where the assets is going to live
00:07:30and what format. And again, for the sake of this demo, I'm just going to go with the recommended
00:07:36option. And I'm not going to show you the rest of these questions because this isn't meant to be a
00:07:40grill me video. Just understand that if you haven't seen it before, this is the general cadence.
00:07:44It's going to ask you a series of questions, give you potential answers and a recommendation.
00:07:48Very similar to plan mode, just plan mode on steroids. So you can see here,
00:07:51we ended up going through 10 questions on the grill me side, and then we transitioned into the codex
00:07:56portion. Now the codex portion is going to create two markdown files for us. We have the plan.md
00:08:02and then the plan review log. So the plan.md is the source of truth for what we're going to create.
00:08:10This is what our final deliverable is going to be. The plan review log.md, this is where
00:08:16cloud code and codex are going to go at it. Codex is going to take a look at the original plan.md and
00:08:21take a look at the overall thing that cloud code has created. And it is in the plan review log that codex is
00:08:28going to say, Hey, this sucks. This doesn't, et cetera. This also gives us a log of their back and
00:08:33forth through all of the cycles. And at the end of this back and forth with codex and cloud code,
00:08:38we will have an updated plan.md. So plan.md is the final deliverable. That's what everything will be
00:08:46built off of. The plan review log is the back and forth and where the sausage is actually made. Another
00:08:52note during this adversarial review is that while it is headless, we still give codex the session ID.
00:08:59So it's not like it's a completely blank slate on codex's part on like iteration one versus iteration
00:09:05two versus iteration three. It always has memory of the entire back and forth with cloud code. So we
00:09:12can see here in round one, that codex found 11 things that it considered issues. And we can also
00:09:18see that cloud code went ahead and updated the plan.md based on the findings that it accepted and felt
00:09:25were valid. In round two, it found four additional findings. We've gone from 11 down to four. And again,
00:09:31the plan was updated. And here on round three, we see that the verdict is now approved. It's at this
00:09:35point that codex and cloud code are now on the same page. Codex has still flagged a couple things,
00:09:40but they're just three low level knits. So they're non blockers. And that's reiterated here at the end
00:09:45where it's telling it is approved round three of five tells us what the final plan looks like,
00:09:50what the two acts bought us and specifically in terms of act two, which is round one and round
00:09:56two of codex and cloud code going at it. You know, we caught real security and correctness holes.
00:10:01There was unbounded client skill slug, case sensitive dedupe bypass, relative email link,
00:10:06raw list bombing vector and a table scanning rate limit. And in the second round, it caught the false
00:10:12fixes. So round one codex said, Hey, here's the issues. Cloud code tried to fix them. And in the
00:10:18second iteration codex is like, those aren't real fixes, right? So it noticed that the double opt-in
00:10:24claimed, but wasn't wired the expression index dedupe that super base JS can't target
00:10:30and the away before response that still blocked unlock was moved to after. So just three rounds,
00:10:38but this is a great time saver versus trying to execute the first plan. Cloud code came up with
00:10:44and then going through the whole troubleshooting process. At the end, it also brings up some open
00:10:49items, mainly like the SQL migration and all that. But that's also cloud code being lazy because it can
00:10:54do that on its own. So back on the website up top, we have the free skill. I click on it. Now it's
00:10:58asking me for my email. And cool. Now I have the skill here that I can download in a .zip file.
00:11:08Obviously in reality, what would I actually want to do? Well, I would probably want the text and
00:11:12everything to actually match the rest of the website, but you can see it created what we set out to do.
00:11:18The point of this video wasn't the specific demo, but just to show you this skill in action. As for
00:11:23how to get these skills yourself, I'll put them down in the pen and comment to make it easy for you.
00:11:27But besides that, that's pretty much all I got. Obviously things you need to know for this is,
00:11:31hey, we're using codex. So you are going to need an open AI account. You're going to need codex
00:11:35downloaded, which is relatively simple to do. And there's no reason you need anything beyond the
00:11:39$20 a month open AI plan to get a lot out of this. This system we've created is also something
00:11:45you could easily swap out for some sort of local model. So if you're like, hey, I don't want to
00:11:50pay open AI $20 a month. I'd rather use something like DeepSeq or whatever, any local or cheaper model
00:11:55you have, really easy to do. Like the bones are there. I would just take the skill I've created,
00:12:00bring inside a cloud code and say, hey, can we swap out codex for insert whatever model you're trying to
00:12:07use? It's really that easy. It's very, very flexible. So there's a lot you can do with
00:12:12it. And I think the bones of it make a lot of sense for those of us who don't consider ourselves
00:12:16expert coders who can take a look quickly and efficiently at what cloud code has done and say,
00:12:22this makes sense. This doesn't. It's just not in a lot of people's wheelhouses,
00:12:26nor does it need to be. Frankly, we have tools that can do this for us. So as always,
00:12:32let me know what you thought. Make sure to check out Chase AI Plus if you want to get your hands on
00:12:35on the Cloud Code Masterclass,
00:12:37and I'll see you around.
Community Posts
No posts yet. Be the first to write about this video!
Write about this video