STOP Using Claude Code OR Codex

CChase AI
컴퓨터/소프트웨어창업/스타트업AI/미래기술

Transcript

00:00:00You are hamstringing yourself if you are trying to choose
00:00:03between Clod Code or Codex.
00:00:05Now, Clod Code has owned the AI discourse for months now,
00:00:08and that's because the gap between Clod Code
00:00:10and the number two option was extremely large.
00:00:14But Codex has quietly closed that gap.
00:00:17GPT 5.5 is an amazing model
00:00:20and is arguably better than Opus 4.7.
00:00:23The usage limits are way more generous
00:00:26with OpenAI's Pro plan than with Anthrovic's Max plans.
00:00:29And yes, that's still the case
00:00:31even though they doubled the five-hour limits.
00:00:34They certainly didn't double the weekly limits, by the way.
00:00:36And the Codex desktop app is a legitimately good product.
00:00:41Now, that's not to say that Codex is better than Clod Code.
00:00:44It's to say that you now have options.
00:00:47And the best play isn't to sit here
00:00:48and try to choose which one of these two good options
00:00:51is better, the best play is to use both.
00:00:54And luckily for us, getting the best of both worlds
00:00:57is extremely easy to do.
00:00:58It takes literally seconds to set up the Codex desktop app
00:01:02with the Clod Code terminal running inside of it.
00:01:05And mastering both tools is also very easy
00:01:07because the Venn diagram of Codex and Clod Code
00:01:11is basically a circle.
00:01:12There's like 99% overlap.
00:01:14So if you learn how to use one of these,
00:01:16you can very easily learn how to use the other.
00:01:18So today I'm gonna tell you
00:01:20what you should be thinking about
00:01:21if you're a Clod Code user
00:01:22who's trying to dip your toes into the Codex waters.
00:01:25We're gonna do a quick demo
00:01:27where I show you how to use these two tools in tandem
00:01:30and we'll have a deeper discussion
00:01:31about why I think you need to be tool agnostic,
00:01:34why we shouldn't pigeonhole ourselves into one coding agent,
00:01:38into one company's ecosystem.
00:01:40Because let's be honest,
00:01:42you owe these companies zero loyalty.
00:01:45So today we're gonna focus on the Codex desktop app.
00:01:47Now there is a Codex CLI, but in my experience,
00:01:51I found that to get the best of both worlds,
00:01:53it's easiest if we use the Codex desktop app
00:01:56with the Clod Code in the terminal inside of it,
00:01:59because you can have a terminal open inside of this app.
00:02:02And the desktop app honestly
00:02:04has some really nice quality of life stuff
00:02:05that I will show you, things like an in-app browser
00:02:07and that sort of thing.
00:02:08So to use it, you just have to go to openai.com/codex
00:02:12and the installer takes like two seconds.
00:02:14Now let's talk about pricing really quick.
00:02:16By and large, if we compare this to anthropic stuff,
00:02:18you're getting more bang for your buck across the board.
00:02:21So it's hard to do some sort of like one-to-one thing
00:02:24because token costs are different.
00:02:26GPT 5.5, if we're talking about cost per million tokens,
00:02:29it's about the same or slightly more expensive actually
00:02:32than Opus, but it uses less tokens.
00:02:34And then usage depends on the time of day or sort of.
00:02:37There's a lot of factors.
00:02:38So it's not like a one-for-one examination
00:02:41or comparison we can do, but big picture,
00:02:44you get more with OpenAI.
00:02:46Now, things you need to note.
00:02:50There's GPT 5.5 and there's GPT 5.5 Pro.
00:02:54GPT 5.5 Pro is only available
00:02:57if you are on the $100 or $200 Pro plan.
00:03:00If you're on the $20 or less plan,
00:03:02you can have 5.5 just sort of straight up.
00:03:055.5 just straight up is still good.
00:03:075.5 Pro is obviously a little bit of a step above
00:03:09and it's a model that actually beats out mythos
00:03:12in some benchmarks.
00:03:14But if you're someone who is coming from cloud coding,
00:03:16like, all right, I'm already doing 200 bucks a month
00:03:18in cloud code, like, do I need to do 100 bucks in Pro
00:03:21to get the full power?
00:03:22I would suggest just start with the 20 bucks a month
00:03:25and get your feet wet and see how you like it.
00:03:27There's no, and if you really like it,
00:03:28you can always upgrade to 100.
00:03:30For me, in my situation, I'm on the $100 Pro
00:03:33as well as the max plans with Anthropic.
00:03:36Once you install codecs, you'll open it up
00:03:37and you will see something like this.
00:03:39Now, before we do the quick rundown
00:03:41of what you need to be thinking about,
00:03:43a quick word from today's sponsor, me.
00:03:46As you know, I recently released a cloud code masterclass,
00:03:48which is the quickest way to go from zero to AI dev,
00:03:51especially if you don't come from a technical background.
00:03:54But I also just released yesterday
00:03:56a codecs masterclass alongside it.
00:03:59And it's for two types of people.
00:04:01It's for those who are brand new,
00:04:02never done any sort of coding stuff,
00:04:04but wanna get into codecs.
00:04:05And it's also for those of you
00:04:07who are a little more experienced,
00:04:09you've been inside of cloud coding,
00:04:10trying to figure out, hey, how do I make the transition?
00:04:12And really, how do I use these two tools in tandem?
00:04:15So it's pretty much everything we talk about today times 10.
00:04:19So if you wanna get your hands on that,
00:04:20as well as stuff like my agentic OS system,
00:04:23OS system, you can find that inside of Chase AI Plus,
00:04:27there's a link to that in the pinned comment.
00:04:29So this is gonna be the very quick
00:04:30five-minute run through of codecs,
00:04:32what you need to be thinking about
00:04:33and some sort of differences.
00:04:35In terms of the UI, very intuitive, honestly.
00:04:37So very chat GPT coded, right?
00:04:40We have the prompt window.
00:04:42I can add photos and files.
00:04:43I can do plan mode right here, it's just a toggle.
00:04:46We have the permissions set up,
00:04:47very similar to permissions inside of cloud code,
00:04:50where we have like bypass permissions, auto,
00:04:52that sort of thing.
00:04:53I can choose the intelligence, AKA the effort,
00:04:55as well as the model right here.
00:04:57I can also very quickly see where I am
00:04:59in terms of what folders I'm operating in.
00:05:01They call them projects.
00:05:02I can work locally or in the cloud.
00:05:05I can do different work trees, that sort of thing.
00:05:07So pretty easy to navigate.
00:05:09Now let's quickly go through the settings tab.
00:05:11So you have general, the work mode.
00:05:13You're gonna wanna be on four coding.
00:05:15This is just gonna give you more technical detail.
00:05:17Permissions shows up again.
00:05:19This is just saying, "Hey, do you even want these
00:05:20as options to be shown to you?"
00:05:22The answer is yes,
00:05:23because you're gonna wanna sit on full access.
00:05:25And then over here in general,
00:05:26most of this just has to do
00:05:27with the environment setup itself.
00:05:30One thing you might notice is right here,
00:05:31follow-up behavior, queue versus steer.
00:05:34We'll talk about that more in detail later.
00:05:36Just keep it on queue for now.
00:05:38Appearance is exactly what you would think it would be,
00:05:40but down here you have pets,
00:05:42which sound kind of stupid at first,
00:05:44but honestly pretty useful
00:05:46because they're basically like a visual hook
00:05:48that lets you know if codecs is working in the background
00:05:51or if it's ready for you to do something else.
00:05:53So it's just like this little thing, right?
00:05:56It goes anywhere on your computer.
00:05:58It sits on top of whatever program you're using.
00:06:00So if I close out codecs,
00:06:03I can still see my guy.
00:06:06And you'll see it later when we're actually doing a task.
00:06:08It will have a little text stream
00:06:09so you can see what it's working on.
00:06:10And then it kind of just goes flat like this
00:06:12when it's done working.
00:06:13So honestly, I love notification stuff like this.
00:06:17Like with Claude code, I just have, you know,
00:06:19an audio hook go off every time it finishes a task
00:06:23because I probably lose more time with agent decoding
00:06:27from just like not getting back to the task
00:06:29after I tell it to do something
00:06:30and then I tab out or walk away.
00:06:32So, you know, hey, use it or don't, doesn't really matter.
00:06:35Then you have configuration.
00:06:37They have some stuff with hooks.
00:06:38It looks like I need to update.
00:06:39Over here is sort of the approval policy
00:06:41and sandbox setting.
00:06:42So this is similar to permissions except on a global level.
00:06:45And then over here for a workspace dependencies,
00:06:48you will want to have codecs dependencies switched on,
00:06:51which it should be by default.
00:06:52Then we have personalization.
00:06:53So you can kind of choose your personality.
00:06:54This is not, this is not agents.md, aka claud.md.
00:06:59So codecs has its own version of claud.md.
00:07:02It's called agents.md.
00:07:03And again, we'll talk about that more detail in a little bit.
00:07:06Personalization's kind of similar, but not exactly.
00:07:09It's more like, hey, I always want you to call me
00:07:11by this name or something like that.
00:07:13There's also memory.
00:07:14This is similar to memory inside of claud code,
00:07:17which is also on by default.
00:07:18I turn this stuff off.
00:07:19It's like, hey, if I tell codecs,
00:07:22hey, I always go to the gym on Tuesdays
00:07:24and then Tuesday rolls around and I say something like,
00:07:26I don't know what to do today.
00:07:27You'll say, oh yeah, you go to the gym on Tuesdays.
00:07:30I don't really care for that stuff to be honest,
00:07:33but up to you.
00:07:34Then the rest of this stuff is like MCP servers,
00:07:37Git, environments, works, trees.
00:07:39This kind of will depend like,
00:07:40depends how technical you are,
00:07:42how deep you want to get into that.
00:07:43Then there's stuff like browser use and computer use.
00:07:46So computer use, you need to be on Mac
00:07:50and then browser use is exactly what it sounds like.
00:07:54Then we also have archive chats and usage.
00:07:56So not too much you have to play around with here.
00:07:59Mainly you're going to be in general
00:08:01and then appearance and configuration.
00:08:03Up top, we got plugins.
00:08:04So Codex has plugins and skills similar to Claude code.
00:08:08The line between those two is pretty blurred.
00:08:11So plugins by and large are almost like skill packs
00:08:14or MCPs that come from providers themselves
00:08:16that you can easily install.
00:08:17So something like Superbase installs a Superbase MCP
00:08:21and the requisite skills.
00:08:22So if I open up a chat now and said,
00:08:24hey, open up or create a database inside a Superbase for me,
00:08:27it just does it.
00:08:28So same thing with all of these and it includes stuff like
00:08:31Chrome and spreadsheets and presentations.
00:08:33And it's a one-click install.
00:08:35Then we have skills works pretty much the same as Claude code.
00:08:38If you just opened Codex,
00:08:40you probably get a pop-up that's going to say something like,
00:08:42hey, we noticed you have all of these skills
00:08:45from another coding agent.
00:08:46Would you like to import them?
00:08:48So it will import with one click of the button,
00:08:51pretty much everything from Claude code
00:08:53or something like open code.
00:08:55So it's able to recognize that on your computer.
00:08:56So that's another thing that makes it really easy to switch
00:08:59between these tools.
00:09:00So I'm like, oh man,
00:09:01I built like this skill army on Claude code.
00:09:03I can't leave it.
00:09:04Well, it's like, no, actually you can.
00:09:06It just automatically throws it in here.
00:09:08And so to use these, you know, you just click on them.
00:09:11You can uninstall them.
00:09:12You can turn them on or off.
00:09:13So again, pretty intuitive.
00:09:15You can also manage them up here,
00:09:17create a skill very easily.
00:09:19And it has its own skill creator skill as well.
00:09:22There's also the automations tab,
00:09:23similar to routines in Claude code.
00:09:25They have some default ones in here.
00:09:27We can go here and automatically create a new automation.
00:09:30You can set it up on the work tree or local.
00:09:32You can put it in a specific project time, all that stuff.
00:09:36You also have the ability to,
00:09:37just like you would in the terminal with Claude code,
00:09:39just say, hey, let's create an automation using X, Y, and Z.
00:09:42And we'll automatically put it in there, but very simple,
00:09:45very intuitive to visually click around here.
00:09:47Now, in terms of navigating the file structure
00:09:50and the space on your computer,
00:09:52the way it breaks it down is projects in chats.
00:09:54So right now I can go inside a project called audit flow,
00:09:58which is when I was on earlier today,
00:10:00or I can add a new project, or I can just say,
00:10:02I click here and I'm in a new chat.
00:10:04So a new chat, it isn't really in any specific folder.
00:10:07This is just like being in the chat window
00:10:09inside of like Claude code desktop.
00:10:11Like I'm just talking to it like it's chat GPT.
00:10:14If I want to work in a specific folder
00:10:16or I want to start a new project,
00:10:17we're going to go to projects.
00:10:19So to do that very simple, you can click up here.
00:10:22You can start from scratch and we'll create a new folder
00:10:24inside whatever you've set as default.
00:10:26I usually just do use an existing folder
00:10:28so I can get a little bit more specific
00:10:31about where I want to go.
00:10:33So in here, we'll do new folder
00:10:36and we'll just call it like YouTube demo codex.
00:10:40And then, yep, hey, do you want to import some settings?
00:10:43Sure, let's do that.
00:10:44Supporting some recent settings changes I had in Claude code.
00:10:47And now you can see here, I'm inside my YouTube demo codex.
00:10:52We're working locally on the main branch.
00:10:54You can all see that over here in projects.
00:10:56So I can say, hey, what's up?
00:10:59And then you now see that chat down here.
00:11:05Now this chat is pretty much the same
00:11:07as having a terminal window open
00:11:08because I can stay in the same project.
00:11:10And if I go up here and I just do start new chat, hi again,
00:11:15I now have two chat windows open,
00:11:18which is virtually the same exact thing
00:11:20as me having two terminals open, right?
00:11:24Same sort of process, open in the same folder,
00:11:27doing their own thing,
00:11:28but still kind of working on the same project.
00:11:29They can see everything between one another,
00:11:31but it's very easy to kind of keep track of it
00:11:34inside of this UI.
00:11:35I can also click on any of the chats.
00:11:36I can copy them, fork it into local,
00:11:38fork it into new work tree, rename, pin it,
00:11:40whatever I want to do.
00:11:41I can also very easily click on the project
00:11:43at three dots right here, open it and explore.
00:11:46So, you know, actually navigating the chats
00:11:49and navigating your file system and having a mental model
00:11:51of where everything is sitting on your machine,
00:11:53very easy to do.
00:11:54And frankly, that's pretty much the codecs desktop app.
00:11:58A lot of other cool stuff going on here, right?
00:12:00You can see like the branch details,
00:12:01very easy to do stuff like get actions,
00:12:03but like that was pretty much the bulk of it, right?
00:12:05Everything I just told you, you can use that.
00:12:08You can build whatever you want.
00:12:09Now we talked a little bit earlier about,
00:12:11hey, you can use the terminal inside of here,
00:12:13up here on the top, right?
00:12:14Toggle terminal, boom, here's the terminal.
00:12:17It's inside my YT demo codecs project.
00:12:20And then we can just run clod, boom.
00:12:25I now have clod code and codecs open in the same project.
00:12:28Now, in terms of bouncing them off one another,
00:12:30couple of ways you can do that.
00:12:31One, we can, and what we'll do right now
00:12:33is we'll have it create some sort of little web app for us.
00:12:36And I can have a plan in codecs, take the plan,
00:12:38copy into clod code, see what it says,
00:12:40copy and paste back and forth through that sort of thing.
00:12:43Or I can have codecs build something,
00:12:44have clod code actually look at the code
00:12:46because they're inside the same directory,
00:12:48figure out what it says.
00:12:50I'm sure there's actually much more sophisticated
00:12:51and simple ways to do it than even this,
00:12:53where you can sort of set up something automatically.
00:12:55I haven't messed with that too much.
00:12:56Point is like the infrastructure's here, really easy to do.
00:13:01We have the best of both worlds.
00:13:03So let's do a simple little demo
00:13:05to kind of put it through the paces.
00:13:07We'll ask it to create sort of a content/research,
00:13:12ideation type web app.
00:13:16One part needs to be able to like pull information
00:13:18from a bunch of sources and give us like possible ideas.
00:13:21Second part, I want it to like be able to synthesize
00:13:24all the information it grabs and come up with content ideas.
00:13:27And then third part,
00:13:29let's have it create some sort of little scheduler
00:13:30at the bottom, maybe like a mini Kanban board
00:13:32to like keep our ideas in track.
00:13:34So it needs to be able to research, ideate,
00:13:36and then actually organize all this data.
00:13:38So let's see what it does.
00:13:39We'll start with codecs.
00:13:40So put it in plan mode.
00:13:42Also in terms of invoking skills and things of that nature,
00:13:47pretty much the same.
00:13:48You can do forward slash,
00:13:50and you can like call a certain skill.
00:13:53So if I was like front end design skill,
00:13:55boom, there we go.
00:13:56Or I can also do at.
00:13:58So I could do at like spreadsheets.
00:14:00And so now it's using the spreadsheet plugin.
00:14:03I can also just use natural language.
00:14:05And just like with Claude code, it should pick it up,
00:14:07but doing like slash commands and at commands,
00:14:10that's sort of how you point at different things.
00:14:11And same thing for pointing at specific files or folders.
00:14:15It also works the same way.
00:14:17And one other thing, context, something to note,
00:14:20the 5.5 Pro has a 258K context window
00:14:25versus Claude codes one million.
00:14:28My take, not really a bad thing.
00:14:31'Cause most people have no idea
00:14:33how to manage their own context.
00:14:35They live in context raw hell.
00:14:37And a 258K pretty much makes it impossible
00:14:40to do that for very long.
00:14:41Now it has auto compaction when you hit 258K
00:14:44and auto compaction has its own slew of issues,
00:14:47especially when we begin compacting the same conversation
00:14:49over and over and over.
00:14:50But like I just showed you,
00:14:52doing the equivalent of forward slash clear
00:14:54is literally just starting a new chat, right?
00:14:59'Cause I pretty much just opened up a new session.
00:15:01So context is one little difference.
00:15:04Let's go ahead and give it a problem and see what it says.
00:15:07So I wanna create a web app
00:15:10that sort of does three things.
00:15:13Ideally you can do it all on the same page.
00:15:15On one hand, I want it to be able to look at AI news
00:15:18over the last 24 hours across the major web sources
00:15:22as well as stuff like YouTube or Twitter.
00:15:25And then I want it to consolidate it into a report.
00:15:29Number two, I want it to be able to take all that information
00:15:31that come up with potential content ideas for me.
00:15:33Like what would a title be?
00:15:34What would the general outline be?
00:15:36It can be just kind of like bullet point format
00:15:38as well as like a hook.
00:15:39And then lastly, I think of it at some sort of like scheduler,
00:15:42maybe like a mini Kanban board
00:15:44where I could then be like,
00:15:46okay, like let's take that idea you came up with
00:15:48and let's do it today.
00:15:49And next idea we can do tomorrow, something like that.
00:15:52So let's kind of walk through that and plan it.
00:15:54And so now we'll go through its plan mode.
00:15:55And the plan mode is basically the exact same as Cloud Code.
00:15:57It's gonna think about it.
00:15:58It's gonna ask you a series of questions.
00:16:01I've noticed with 5.5 Pro on Extra High,
00:16:05it tends to ask quite a few questions,
00:16:08although maybe it was just the projects I was working on.
00:16:10And in terms of speed,
00:16:13it's a little,
00:16:15I think it's probably a little bit slower than Opus.
00:16:18Although, you know, I don't have like hard numbers on that.
00:16:21It's just kind of what the vibes have been.
00:16:22At the same time,
00:16:23if I'm more just doing like a back and forth chat,
00:16:255.5 feels a lot snappier than Opus.
00:16:28So if it's doing a bunch of tool calls, a little bit slower,
00:16:31it's just chatting rather fast.
00:16:33So here's the plan Codex came up with.
00:16:35Build a greenfield single user local web app
00:16:38with Next.js, TypeScript and SQLite.
00:16:40The app will have one main dashboard for three flows,
00:16:43click the last 24 hours of AI signal,
00:16:45generate a concise report plus YouTube video ideas
00:16:48and schedule selected ideas on a mini Kanban board.
00:16:50So no paid APIs, created RSS feeds
00:16:54and local Olama generation.
00:16:55So easy way to get Clod Code involved
00:16:58is we're just gonna copy this, paste it into Clod Code,
00:17:00see if it has any other ideas or any blind spots.
00:17:04So I said, Codex came up with this plan for our app.
00:17:06What do you think what's missing?
00:17:08So Clod Code came back.
00:17:09So the plan's solid, but has some gaps,
00:17:12some soft concerns as well as some nitpicking.
00:17:14So what I'm gonna do, come back into Codex,
00:17:17paste that in there and just say,
00:17:18what do you think of this?
00:17:22And submit.
00:17:25Now, we could continue to have this back and forth forever.
00:17:30And for the sake of time, we'll stop here.
00:17:32But the idea is we now have a second set of eyes
00:17:35on what the heck it is AI is coming up with as a plan.
00:17:38And I think this is super important,
00:17:40especially if you're someone who don't,
00:17:41who doesn't come from a technical background, right?
00:17:43Because the problem is you go to AI, you have an idea,
00:17:47it gives you a plan.
00:17:48If you have no idea what right is supposed to look like,
00:17:51you're kind of just like, sick dude, thumbs up, go for it.
00:17:55And it could be missing a whole lot.
00:17:57Now we try to get around that
00:17:58by just like asking more questions, being more thorough,
00:18:01asking things like, what am I not thinking about?
00:18:03What would an expert ask?
00:18:05But what if we do all that and we have Clod Code take a look?
00:18:10And this also implies that the inverse,
00:18:11having Clod Code build it and bring it into codec.
00:18:13So if nothing else, it should kind of just give you
00:18:16that warm and fuzzy feeling inside that like,
00:18:17all right, like this does make sense.
00:18:20Multiple AI experts are telling me it's a solid plan.
00:18:24And codecs even says,
00:18:25I agree with the critique's main diagnosis.
00:18:26The original plan would reliably summarize what happened,
00:18:29but the product you actually describe
00:18:30needs what is worth making a video about today.
00:18:33That requires trend signals, ranking,
00:18:35and competitor saturation checks, not just ingestion.
00:18:38So it is making some changes to the plan
00:18:40because of what Clod Code came back with.
00:18:42And obviously this sort of dual model approach
00:18:44is something you can apply to any part of your project.
00:18:47And so here's a new plan with the updates.
00:18:49And like I said, for time,
00:18:50we're just going to execute it after this first pass.
00:18:5323 minutes, 21 seconds.
00:18:54And it says it has implemented the full local AI trend planner.
00:18:58What's in place, goes through it, key files,
00:19:00verification passed, it created a read me.
00:19:04So I can click on the read me and you can kind of see this,
00:19:06like in app sort of thing here.
00:19:08So you can see what it actually wrote up
00:19:10and then it shows all the different files.
00:19:12So if I click on all the files,
00:19:15it will quickly show sort of like what it created.
00:19:18Obviously it hasn't deleted anything
00:19:19since it's its first pass, but it would show that as well.
00:19:22If I click on any of these files,
00:19:24I can also show it in review.
00:19:26Once it's inside review, there is a like diff viewer.
00:19:29I can do some get actions here and I just,
00:19:33it's really easy to see what it's actually executed.
00:19:36Again, I love the terminal.
00:19:37You probably love the terminal,
00:19:38but the terminal does have some limitations
00:19:40when it comes to the convenience factor
00:19:42of seeing everything all in one place.
00:19:44So before we even have Claude code take a look,
00:19:45let's say spin up the dev server for me
00:19:49and open it up in the sidebar browser.
00:19:54I think I've genuinely gotten like so bad at typing
00:19:57since I've been using AI so much
00:20:00and just like voice dictation over the last year.
00:20:02I actually have completely lost the ability
00:20:04to like type like a single sentence without errors.
00:20:07So what it's gonna do is it's going to spin up the dev server
00:20:11and it's gonna show us the actual webpage here
00:20:14in the in-app browser, which is nice.
00:20:16And so now we can see the webpage in the browser.
00:20:18I'll move over here so it's a little easier for you to see.
00:20:21So here's what it created.
00:20:23We have the AI trend planner.
00:20:26We can run a scan.
00:20:27We can ingest stuff, report ideas.
00:20:29And overall, I think actually for the first pass
00:20:32kind of went with like this, I think it's like a sort
00:20:35of brutalist type approach.
00:20:37I think it looks pretty good.
00:20:39I mean, I kind of like it, I don't know.
00:20:40Everything's AI slop these days, right?
00:20:43But many Kanban, can I drag these?
00:20:46No, can't drag those.
00:20:48Would like to be able to do that.
00:20:50Signal feeds, sources, okay.
00:20:55On the surface, I don't know if any of that actually works,
00:20:58but it looks decent at the beginning.
00:21:02So let's see what happens if I do run full scan,
00:21:05fetching sources.
00:21:08And while it's doing this, what should we do?
00:21:11Well, we should just have Cloud Code take a look
00:21:12at its work.
00:21:13Hey, can you take a look at what codecs built
00:21:17on its first pass for our application?
00:21:21Any glaring weaknesses, anything you would change
00:21:25is from my understanding, everything should actually be wired
00:21:28up properly and working, but do you see any issues
00:21:31that somehow slip through the cracks?
00:21:34Well, it's saying Ollama didn't even work.
00:21:36So we probably have to figure out something with Ollama
00:21:38on that end, but also now ago.
00:21:42Overall looks pretty cool.
00:21:42And like obviously we can also go onto this
00:21:44on our local browser.
00:21:46One of the things we can do here and kind of reminiscent
00:21:48of things like cloud design is you can like annotate things
00:21:52or leave comments.
00:21:53So I could highlight this, leave a comment
00:21:58and say something like, can we make this italic, italicized?
00:22:03And that then puts it in a annotation right here
00:22:10and I can add additional follow-up changes
00:22:13or I could just send that right now.
00:22:15And then you kind of have the ability
00:22:17to annotate anything you want.
00:22:20You can also quickly take a screenshot.
00:22:22I do that and then I could paste it in there.
00:22:24So it makes also sort of front-end design reviews
00:22:28and iterations like that a lot easier.
00:22:30And hey, there, we now have the italics.
00:22:33Oh, and Codex did tell me at the beginning,
00:22:35but I was lazy and didn't even read it,
00:22:36was hey, to use the local AI generation, run Ollama poll,
00:22:40Ollama 3.18 or set Ollama model to a model you have.
00:22:43I have a few models on my computer.
00:22:45So we'll just say, hey, can you just find it?
00:22:47Hey, so I'm pretty sure I have a couple Ollama models
00:22:51already on my machine.
00:22:52Can you take a look at which ones those are
00:22:55and then properly wire them up?
00:22:57Also, let's check to see if these links actually work.
00:23:00So it says AI Slop is killing online communities
00:23:03from Y Combinator.
00:23:05So let's copy that.
00:23:06Yeah, no, it's a real thing.
00:23:08Very cool.
00:23:09So down here, Claude came back with its review.
00:23:12So it said it found several real bugs,
00:23:15came back with 20 new bugs.
00:23:19And then it says, bottom line, wiring is correct.
00:23:22The pipeline flows end-to-end, bugs that will surface fast,
00:23:27and then has like some timestamp issues,
00:23:30competitor self-warning and some other stuff it brought up.
00:23:32So, you know, pretty good.
00:23:34I mean, 20 things.
00:23:36I also wonder if telling Claude code, like Codex wrote this,
00:23:39if it becomes a little extra adversarial,
00:23:42which I would love.
00:23:43Probably can actually like make that a skill.
00:23:45Considering there is a skill for the Codex plugin
00:23:48inside of Claude code, which is literally called
00:23:50Adversarial Review.
00:23:51So Codex went ahead and realized I have GLM 4.7 flash
00:23:55on my machine.
00:23:56It wired it up and ran the trend report again.
00:23:59So you can see that here.
00:24:00So this is an actual legit report
00:24:02based on everything it grabbed.
00:24:04So I can look at these video ideas.
00:24:05So why AI Slop is killing online communities
00:24:07if I hit tomorrow.
00:24:09Let's see if it actually puts it down there.
00:24:15Doesn't look like it's kind of spazzing out.
00:24:18So let's have Claude code to fix that.
00:24:21Hey, so when I click on one of the video ideas,
00:24:24for example, the why AI Slop is killing online communities.
00:24:28If I click on tomorrow, I kind of just get a loading bar
00:24:32and nothing actually happens.
00:24:33Can we fix that?
00:24:35And also secondly, right now on the Kanban board,
00:24:37I'm not able to actually move things around on the board.
00:24:41Once they're in a specific slot, like inbox,
00:24:43I don't have the ability to move them to today or tomorrow,
00:24:45whatever.
00:24:46And so I think this is how you would sort of do this
00:24:48back and forth.
00:24:50Like you can see here,
00:24:50you can have both of them kind of work on something
00:24:52at the same time.
00:24:53You can have something go to Claude code
00:24:55if you feel like codecs didn't hack it.
00:24:57And then I think you can have them sort of like make up
00:24:59for certain weaknesses with one another.
00:25:01I mean, usually I would say Claude code
00:25:03tends to do a little better on the front end design
00:25:05side of things and design in general than codecs.
00:25:08But to be honest, I kind of like how it came out
00:25:11on the first pass.
00:25:12So, you know, I could do a whole demo showing you,
00:25:15hey, Claude, go do, you know, work on the front end design.
00:25:17And there's probably a little bit we could work on here,
00:25:19but I think it does pretty good.
00:25:21And what I really wanted you to see from this demo
00:25:23was just like how easy it is to set this up.
00:25:26And the fact that it saw things like Claude code,
00:25:29it saw 20 potential bugs that codecs didn't pick up on,
00:25:34on its first pass.
00:25:35And I think the sort of compound interest of having them
00:25:38check each other over and over and over again,
00:25:41throughout a project kind of pays for itself over time
00:25:45because at the beginning, like actually we were using
00:25:47a bunch of tokens to do this, but in the long run,
00:25:49if we're able to grab these bugs early,
00:25:52pick out these weak points right away,
00:25:54I think in the aggregate, you actually save tokens.
00:25:57And obviously we can always open this up
00:25:59in our normal browser as well.
00:26:01It looks like it was able to fix the actual navigation
00:26:06on the Kanban board.
00:26:07Now, the last thing I want to talk about very quickly
00:26:09was sort of the bigger idea of being tool agnostic,
00:26:12because I think a lot of people,
00:26:14they think that, oh, like I am tool agnostic.
00:26:16Like I'll switch to the best tool tomorrow
00:26:18if there's something better than Opus,
00:26:19or I will switch to the other best tool
00:26:21if there's something better than GPT 5.5.
00:26:23People think that, and in reality,
00:26:25they don't actually do that.
00:26:26In reality, what happens is you get used to one tool,
00:26:29you get ingrained in the habit of using it.
00:26:31And then for a lot of people,
00:26:33you become like weirdly tribal about it,
00:26:35where it's like, no, like I'm an anthropic guy,
00:26:37I'm a collage code guy, I hate open AI,
00:26:40I hate Sam Altman or the complete reverse,
00:26:43as if like they're a sports team.
00:26:46You shouldn't care.
00:26:48You really shouldn't care.
00:26:49You should be very willing to switch all the time.
00:26:51And it's really easy to switch
00:26:52if you consistently use all of them and use them in tandem,
00:26:56because I think we're getting to the point,
00:26:58I think it's only going to get worse where it's like,
00:27:00which model is best isn't so obvious,
00:27:03because, hey, half of us don't even buy
00:27:06what they're selling in terms of the benchmarks
00:27:08and B, like they're all kind of converging
00:27:09to be like really, really good.
00:27:11And the better they get,
00:27:13the more they're kind of blowing past
00:27:15what the average person is doing.
00:27:17The average project that 99% of people are doing
00:27:19can be done by the models that exist today.
00:27:21So what are we going to do
00:27:22with the models that exist five years from now?
00:27:25So I feel like you're kind of getting further ahead
00:27:29in the long run if you pit them against one another,
00:27:31instead of being like, I'm a cloud code guy,
00:27:33or I'm a codex guy.
00:27:34And luckily for us, they haven't created
00:27:36these like walled gardens
00:27:37where I can interact with both simultaneously.
00:27:39If anything, they've made it really easy
00:27:40to interact with them,
00:27:42like codecs creating the cloud code plugin
00:27:44and the ability to like bring over skill files
00:27:46and that sort of thing.
00:27:47So I think we're kind of in a great age for AI
00:27:51as much as people kind of like to be doomers about it
00:27:53and say, oh, like the prices are increasing.
00:27:54I think in reality, we're in a great spot
00:27:56and it's only getting better
00:27:57and you can make it better for yourself
00:27:59if you use all the tools available.
00:28:02So that's where I'm going to leave you guys for today.
00:28:03I hope you were able to get something out of this.
00:28:05As always, let me know what you thought.
00:28:07Make sure to check out Chase AI+
00:28:08if you want to get your hands on my cloud code
00:28:10and codecs masterclass.
00:28:12And besides that, I'll see you around.

Key Takeaway

Integrating Claude Code directly into the Codex desktop terminal enables a dual-model workflow that leverages GPT 5.5's token efficiency and Anthropic Opus's superior frontend design logic to identify bugs early in the development cycle.

Highlights

  • OpenAI Codex desktop app includes an in-app browser and a sidebar for real-time frontend design reviews with annotation and screenshot capabilities.

  • GPT 5.5 Pro features a 258K context window with auto-compaction, contrasting with the one million token context window of Claude Code.

  • The Codex desktop app automatically detects and imports skill files and MCP servers from other coding agents like Claude Code with a one-click installer.

  • Running Claude Code within the Codex desktop terminal allows for an adversarial review process that can identify up to 20 potential bugs in a single pass.

  • GPT 5.5 Pro is restricted to users on $100 or $200 monthly Pro plans, while standard GPT 5.5 is available on $20 monthly plans.

Timeline

The Strategic Advantage of Tool Agnosticism

  • Relying on a single AI ecosystem creates artificial limitations in development speed and quality.
  • OpenAI Codex has closed the performance gap between GPT 5.5 and Anthropic Opus 4.7.
  • Loyalty to a specific AI provider offers zero functional benefit to the end user.

The current AI landscape has reached a point of convergence where multiple models offer high-level coding capabilities. Choosing between them is a false dilemma because the most effective strategy involves using both simultaneously. Learning one tool translates to the other due to a 99% overlap in functional logic and command structures. The Codex desktop app serves as a centralized hub for this hybrid approach by hosting other terminal-based agents.

Pricing Structures and Model Access

  • OpenAI Pro plans generally offer more generous usage limits compared to Anthropic Max plans.
  • GPT 5.5 Pro is exclusive to high-tier $100 and $200 subscription levels.
  • Token consumption for GPT 5.5 is often lower than Opus despite a similar cost per million tokens.

Cost comparisons between models are complex because efficiency varies by task, but OpenAI often provides higher volume for the price. Users transitioning from a $200 Claude Max plan should start with the $20 Codex entry point to test compatibility before upgrading to the $100 Pro tier. GPT 5.5 Pro is noted for outperforming Mythos and other benchmarks in specific technical coding tasks.

Codex Interface and Environment Configuration

  • Plan mode and permission toggles mirror the functionality found in Claude Code.
  • Desktop 'pets' provide a persistent visual indicator of background agent activity across all open programs.
  • Follow-up behaviors allow users to choose between queuing commands or active steering.

The Codex UI is designed for intuitive navigation with specific modes for technical coding and general chat. The appearance settings include functional UI elements like 'pets' that change state to signify when a task is complete or if the agent is idling. Users can also configure global approval policies and sandbox settings to manage how the agent interacts with the local file system.

Skill Migration and Plugin Integration

  • Codex plugins function as one-click MCP skill packs from specific service providers.
  • The application can automatically scan the computer to import existing skill libraries from other agents.
  • Project-based chat organization allows for multiple simultaneous terminal sessions in the same directory.

One-click installs for services like Supabase enable immediate database creation through natural language. This system removes the barrier of switching tools because users do not have to rebuild their 'skill armies' from scratch. The interface also supports forking chats into new work trees or local branches to test different development paths safely.

Dual-Agent Workflow Execution

  • Running Claude Code inside the Codex terminal allows for real-time cross-model plan verification.
  • Context management is handled via auto-compaction in GPT 5.5 Pro or by starting fresh chat sessions.
  • Using a second model as a 'second set of eyes' identifies blind spots in technical architectures.

Integrating both agents allows for a collaborative planning phase where one model critiques the other's proposal. In a demo for an AI trend planner app, Codex proposed a stack using Next.js and SQLite, which Claude Code then audited for missing features like trend signals and competitor saturation checks. This back-and-forth ensures that the final development plan is robust and vetted by different architectural logics.

Frontend Review and Final Implementation

  • In-app browsers enable direct manipulation and annotation of the live dev server.
  • Adversarial review skills in Claude Code can uncover up to 20 bugs missed by the primary coding agent.
  • Local model integration via Ollama allows the app to process data without external API calls.

The final stages of development benefit from the Codex sidebar browser, where users can highlight UI elements to request specific styling changes like italics. While Codex handles the bulk of the initial code generation, Claude Code is used to fix functional errors in the Kanban board navigation and wire up local Ollama models. This compound approach results in a functional application that captures real-time data from sources like Y Combinator and Twitter.

Community Posts

View all posts