You Can Make Claude + Codex Plan Together. Here's How.

Hinzugefügt: 2026-05-05

[00:00:00]Everyone loves building way more than planning, but the truth is the more time you spend specking and planning, the less time you spend going back and forth with something like Cloud Code. And I don't use things like the / Ultra plan because it becomes very/ ultra expensive. But if you do want to bring your Robin to your Claude Code Batman, then you can use what I'll show you in this video, which is called /Claudeex.

[00:00:22]And this is basically a loop where you can either start with a cloud code plan that you audit in multiple rounds with codecs or you can have them interchangeably work together. So in this video I'm going to walk you through how it works and by the end I'm going to actually give it to you. So if that sounds like a plan then let's dive in.

[00:00:38]In my last video I walked you through how to use the native OpenAI plugin to do something like /codeex and use any one of these various commands here. And while this is awesome, it doesn't give you the ability to automate a workflow where you can actually put this dynamic duo together and have them plan in unison. You're still technically driving every single move. So the whole point of/cloudex is this is meant to change all of that. All you have to do is do plan and then we do a space here. You can tell it how many rounds of back and forth you want to go through. Now, in this case, I have it start with a clawed codebased plan, and then I have Codeex go over the plan three different times.

[00:01:15]Depending on what it is that you're trying to put together, you could make this one round, two rounds, or if you have the token budget, you can go for five rounds. So, you could theoretically do dash-3 for three rounds and say something like build me a clone of to-d doist, the to-do list app. And then it would go together and put together a plan from Claude through a plan.mmd file. Then it would have Codeex audit it, look for flaws, patch the flaws, then go back to Claude, see if it's revised it, and then go back to Codeex and look for any additional flaws it didn't get in its first pass. So this is the 50,000 ft view of how it works. The first thing that happens is that Claude drafts a plan, then Codeex reviews said plan and then comes up with its finding and where it thinks it could be better.

[00:01:57]Then Claude reads the findings and integrates it into its original plan.

[00:02:01]And then finally, Claudex asks, did we reach the desired number of rounds to go through revision? On average from testing it so far, it seems that two to three rounds is sufficient to get the 8020 of the planning bugs. So, if you were to take an X-ray machine, this is what /Cloudex would look like behind the scenes. Step one is where you'd add what it is that you're trying to build or whatever you're trying to add on to what already exists. And after you send that off, this will initiate a script. It's called a bash script. This will essentially execute the initialization of the loop. So if you put no rounds, it will default to three different rounds.

[00:02:34]And if you have a prompt, it will take said prompt and then it will add it to what's called a YAML file. This essentially like a notepad or a sticky note that keeps track of how many rounds this will go for, what is the main prompt we're optimizing for, and what is the end state we're going for. And once we have the prompt, this will kick off step number two, which is essentially initializing this script. It's called a bash script. What it's doing behind the scenes is it's taking the prompt and then putting together a file. This file is called a YAML file. This will specify what it is that we're building or doing, how many rounds we've set, and what round are we currently in. Now, step four is the magic piece that lets us all work together. So, Cloud Code has these things built in called hooks. One of them is called a stop hook. This essentially forces claude code to stop and take a look at the state file to make sure it's on track and it knows exactly where we are and how many more rounds it needs to go and invoke the codeex plugin. So you can think of this as a bouncer that tells claude code either okay you can leave or wait we have to do something else. If it decides we have to do something else this is where it's going to invoke the codeex command line interface to go and actually get codeex's opinion on your existing plan. And then last but not least you get all the findings from codeex. This is then fed to cloud code to internalize and see which parts of its existing plan should it edit. Now, I've only shown you the plan function, but I also have a few others. So, I'm going to walk you through exactly what these do in case you want to take my code, edit it to your own specifications, and know exactly how they work. So, when you do slash review, codeex will just audit what you have so far and give you written commentary about it. It won't auto change your plan. It won't edit anything, and it won't tell Claude Code to do anything other than provide an asset. The cancel command acts as an emergency break where if you see it steering off of course or you feel like your initial plan wasn't good enough or wasn't deserving of the token burn it was about to go through then you can just gracefully stop it and restart. And then you have / rollback which is meant to be the equivalent of a nuclear cleanup. So if you've started a plan and you've canceled it and you feel like you want to wipe the slate clean completely you can use this function that's way more involved than the emergency break. So the TLDDR is that the two ways you can use Claudeex is either to create a brand new project from scratch or you can grill an existing plan. And one additional thing that I've done that I haven't shown you in the demo is I also injected an ask user input tool. This is essentially the very mechanism that claude code can provoke a series of multiple choice questions. So if you put a very vague plan, it will ask you for more feedback before it even goes into this loop. And by the way, I'm holding myself back purposefully to not go full nerd mode to make this as accessible as possible. But if you want to actually learn exactly how I built this step by step and learn all of the plumbing involved to do something like this on your own, I have a dedicated brand new module that I've added to our living course all about how to use this and how to build on it. So, if that interests you, then check it out. Now, this is what it looks like in the wild. Imagine you're trying to build a special web page where it's not just a web page. It's a web page that can also track what everyone's doing, record sessions on that web page of where people are clicking, where they're spending their most time. And you didn't want to use a third-party service and pay for that as well. So, this is a more tricky task that I wrote a scope for and architecture for. Let's be honest, I had AI write it for me. And it walks through the diagram of exactly what the scope of this entire build is. Now, what it is is less important. I more want to show you how you would use this. So you do /claudex plan in this case you don't have to specify rounds three it does it by default and I would just say plan the pulse page build using and then I would tag the files at scope MD and then at architecture MD. This will automatically force claude code to read these before it does anything else. And then you can see the flow from here. It will write the plan and this is the claude code version of that plan. And then as you scroll down, you will see it ran three different stop hooks and then it stops it exactly where it needs to. And then as you scroll through, you'll see that the plan is updating. Every time you see a red, that is a deletion from the plan.

[00:06:36]Every time you see a green, that is an addition from the plan. And that addition is provoked by something like codeex. So, as you go through, this will go down endlessly because it goes very deep in terms of checking every single part of the plan, double-checking if it's thoughtful enough and if it's thought of the second order and the third order consequences. And after it takes its 15 to 20 minutes to go back and forth, we get to something like this where it says, "Round three, hard stop reached. All 10 findings by codeex have been actually integrated into the plan."

[00:07:05]And it walks you through all the details. And if we compare the robustness here, we have one tab that goes through the execution just using Claudeex and one that uses only Claw to create the plan. And you'll notice right away without getting into the trenches and the details, you have all five phases here. And it is way more detailed than what I'm about to show you. So this is number one, and this is the one only from Claude. You'll even notice visibly that the phases are a little bit more frail in terms of what they're going to be doing when you go through the details depending on what it is. You could even use this by the way for creating things like assets like PowerPoints, Excel files, docx, whatever you want. The whole point of this is that you have a much richer planning mode so that by the time you actually execute, you won't theoretically oneshot something, but you can increase the likelihood that you will get the same thing done in less steps. And that's pretty much it. So, as promised, I will give this slash command to you so you can take it, break it, change it to your specifications, and use it to get even better planning.

[00:08:05]You'll find the link to the GitHub along with a guide on how to use it in the second link in the description below, completely free. If you want to go infinitely deeper and see exactly how you could have built this for yourself, how you can build on top of it, and in general become a master of cloud code and codecs, then check out the first link in description below, and I'll see you in my early A adopters community.

[00:08:24]For the rest of you, if you enjoyed the video, I'd super appreciate a like and a sub on the channel. And if you feel so pulled to do so, a comment on the video would really help the reach.