17:06Ras Mic · Tutorial
Trust Me, You're Sleeping on Cloud Agents
A walkthrough of the four ways cloud-hosted coding agents replace manual bug verification, QA, and code review — plus the setup behind all of it.
July 15thThe argument in one line.
Codex only becomes a genuine agentic tool once you stack its plugin system, in-app browser, computer-use mode, parallel threads, and goal-plus-feedback-loop prompting on top of it, instead of treating it as a single prompt box.
Read if. Skip if.
READ IF YOU ARE…
- You already use Codex, Claude Code, or a similar AI coding agent and want to get more out of it than single-shot prompting.
- You're deciding between competing agent frameworks or SDKs and want a way to test options against each other instead of guessing.
- You run repetitive computer or browser tasks (form-filling, data scraping, QA passes) that an agent could take off your plate.
- You want a working mental model for when to hand an agent a long, unsupervised task versus when to stay hands-on.
SKIP IF…
- You've never used an AI coding agent and want a from-scratch setup guide, not a power-user's workflow.
- You're looking for a Codex-vs-Claude-Code comparison; this video only covers Codex.
The full version, fast.
Most people use Codex as a prompt box and miss what it can actually do. The video lays out seven habits: install only plugins that match your real workflow (computer use is the one non-negotiable), use the in-app browser to annotate a running app and to have the agent test it, hand repetitive desktop or form-filling work to computer use, spin up parallel threads when you're choosing between competing frameworks or architectures, match effort level to task difficulty instead of maxing it by default, only trust slash goal with a clearly defined finish line plus a measurable feedback loop, and use a review-until-passing skill to keep code review honest. The through-line: agentic value comes from layering these habits, not from a better single prompt.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:44
01 · Cold open: stop using Codex like a prompt box
Direct callout that most viewers underuse Codex, plus the video's promise: seven things that make you a pro user.
00:44 – 03:10
02 · Habit 1: Plugins matched to your workflow
Computer use is the one must-install plugin; everything else should mirror tools already in your workflow (GitHub, Vercel, Convex, Linear), not every app you've ever used.
03:10 – 06:27
03 · Habit 2: In-app browser for annotation and agent QA
A logged-in, persistent browser session inside Codex; stack visual annotations on a running app instead of writing text prompts, and point the agent's browser at your app to have it test and report back.
07:18 – 10:19
04 · Habit 3: Computer use, goofy test and real knowledge work
A folder/knock-knock-joke stunt proves the agent can drive the OS with its own cursor without blocking you, then the real use case: scraping leads and filling out forms as automated knowledge work.
10:19 – 14:21
05 · Habit 4: Thread maxing plus a machine-config aside
Parallel threads compare competing agent frameworks (Vercel AI SDK, Cloudflare Agents SDK, Anthropic SDK, OpenAI SDK, Effect) for a real client use case; a side note on using Codex CLI for home-network and dashboard setup.
14:21 – 15:45
06 · Habit 5: Model selection, Sol and effort levels
Living mostly in GPT 5.6 Sol, switching between medium and extra-high effort based on task difficulty, with high used only rarely.
15:45 – 18:06
07 · Habit 6: Slash goal only works with a feedback loop
A client-site speed project run under slash goal for hours, succeeding because it had a defined finish line (a perfect Lighthouse score) and a way to check its own progress.
18:06 – 20:40
08 · Grep loop skill, T3 Code, and sign-off
Greptile's review-until-5/5 skill as another feedback-loop pattern, a quick T3 Code plug for cross-device work, and the closing subscribe ask.
Lines worth screenshotting.
- Using an agentic coding tool as a plain prompt box wastes most of what it can do; the gap is in workflow habits, not the underlying model.
- The only plugin worth installing by default is computer use, everything else should mirror tools already in your actual workflow.
- An agent's in-app browser can hold a real, logged-in session, so it browses your accounts as you rather than as an anonymous instance.
- Annotating directly on a running app beats writing a text prompt to describe which element you mean, and annotations can be stacked before sending.
- Pointing an agent's browser at your own running app and asking it to use the product turns manual QA into an automated first pass.
- A computer-use agent runs with its own cursor in its own session, so it can work the machine without blocking you from using it at the same time.
- The most durable use of computer use isn't inside a codebase at all, it's repetitive desktop and browser work like scraping a list of sites into a spreadsheet.
- When you can't decide between two competing approaches, give each one its own thread and let both build a working version instead of debating in the abstract.
- The same agent that codes is also strong at pure machine-level setup work, configuring a network, auditing installed software, building local dashboards.
- Effort level should match task difficulty; living in a middle tier most of the time is a deliberate cost and speed tradeoff, not a compromise.
- An autonomous, unsupervised task only works if it has a clearly defined finish line and an objective way for the agent to check its own progress.
- A measurable score, like a Lighthouse audit, makes a strong feedback loop because the agent can read it and improve against it without a human in the loop.
- A review-until-passing loop, trigger review, fix what's flagged, push, re-trigger, repeat, turns a one-time code review into a standard the agent keeps working against.
- Running the same subscription's lighter cross-device interface can replace a second paid tool for picking up a thread away from your main machine.
Seven habits that turn Codex agentic
AI AGENT WORKFLOW
Codex stops being a prompt box once you layer plugins, an in-app browser, computer use, parallel threads, and goal-plus-feedback-loop prompting on top of it.
02Habit 1: Plugins matched to your workflow
- Install plugins that mirror your actual workflow instead of every tool you've ever used; extras just add clutter you'll never open.
- Computer use is the one plugin worth installing by default, since it's what lets the agent act outside the chat window at all.
03Habit 2: In-app browser for annotation and agent QA
- Annotate directly on a running app instead of writing a text prompt about which element you mean, and stack several annotations before sending them.
- Point the agent's browser at your own running app and ask it to use the product; this automates a first-pass QA test before you ever look at it.
04Habit 3: Computer use, goofy test and real knowledge work
- The computer-use agent runs with its own cursor in its own session, so it works the machine without blocking you from using it at the same time.
- Its most durable use case is repetitive desktop and browser work, like scraping a list of sites into a spreadsheet or filling out forms end to end.
05Habit 4: Thread maxing plus a machine-config aside
- When you can't decide between two approaches, give each one its own thread and have both build a working version instead of debating which is better in the abstract.
- The same tool is also strong at pure machine-level setup work, configuring a home network or building local dashboards, separate from any coding task.
06Habit 5: Model selection, Sol and effort levels
- Match effort level to task difficulty instead of maxing it out by default: routine work runs on medium, harder problems get bumped to extra-high.
- Living in a middle effort tier most of the time is a deliberate cost and speed tradeoff, not a compromise.
07Habit 6: Slash goal only works with a feedback loop
- A long, unsupervised task only works if it has a clearly defined finish line and an objective way for the agent to check its own progress against.
- A measurable score, like a Lighthouse audit, makes a good feedback loop because the agent can read it and improve against it without you in the loop.
08Grep loop skill, T3 Code, and sign-off
- A review-until-passing loop, trigger review, fix what's flagged, push, re-trigger, repeat, turns a one-time code review into a standard the agent keeps working against.
Terms worth knowing.
- Computer-use agent
- An AI agent mode that controls the actual operating system, mouse and keyboard, rather than just a chat window, so it can act outside your codebase entirely.
- Thread maxing
- Running the same task across multiple separate agent conversations in parallel, one thread per competing option, so you can compare finished outputs instead of guessing.
- Slash goal
- A Codex mode where you hand the agent one defined objective and let it work autonomously in a loop until that objective is met.
- Feedback loop (agent)
- An objective, machine-checkable signal, like a test score, that lets an autonomous agent verify its own progress without a human reviewing every step.
- Grep loop skill
- A named skill, built by code-review tool Greptile, that repeatedly triggers a code review, fixes what's flagged, and re-submits until the review passes cleanly.
- Effort level
- A setting that controls how much reasoning and compute a model spends per response, traded off against speed and cost.
- T3 Code
- A cross-device coding interface that can run on the same underlying subscription as Codex, useful for picking up a thread from a phone or a different machine.
Things they pointed at.
18:06toolGreptile (grep loop skill)
19:36toolT3 Code
03:42toolWhisper Flow
10:50toolVercel AI SDK
10:53toolCloudflare Agents SDK
Lines you could clip.
00:06
“If you, the person watching this, are just using it as a prompt box, you're missing out.”
01:00
“If you use anything in dark mode, I judge you.”
05:15
“This is basically the QA process automated.”
16:53
“You need a feedback loop, you need a defined goal.”
18:25
“It's like a slop blocker.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphor
The OpenAI Codex app is an amazing GUI slash harness. There's a lot of cool things that you can do with it, but a lot of people are not maximizing it. If you, the person watching this, are just using it as a prompt box, you're missing out.
In today's video, I'm gonna be sharing you how to maximize your codex usage. In my last video, when I shared a little bit about how I use it, a lot of you are begging me, telling me to make this video, this video is here. Now a couple things before we get started.
I'm gonna be sharing with you seven things that are going to make you a pro codex user. If you're a developer, if you're using it for business, whatever the case is, for my devs who are building projects out there, this one's going to be a good one. Another thing, I moved new space.
I missed y'all. Did y'all miss you? Come give me a hug.
Give me a gun. In all seriousness, almost at a 100,000 subscribers.
This is the subscriber count right now. It would mean the world to me if you subscribed. But that's what you're gonna get, a codex video.
I'm gonna show you how I use it, how I maximize it. Sit back, relax. Let's get into it.
So first things first, I use it in light mode. If you use anything in dark mode, I judge you. I think you are a follower and not a leader.
Uh, true leaders use things in light mode. That being said, the first thing we're gonna talk about are plugins. The plugin system in Codex is fantastic.
Now, I'm going to tell you the must have plugin just in case it's not installed. If you do not have the computer use plugin, what are you doing?
Install that now. Now that you have that installed, and we'll talk about computers use in a second. The plugins that you should install are the ones that are part of your workflow.
Whenever I'm building an application, I can think of GitHub, Vercel, Convex, and Linear. These things are a part of my workflow.
So therefore, when using the codecs out, the first thing I'm thinking about are the plugins, the things that I use in my workflow. Are they plugins?
Vercel? Yes. Convex?
Yes. Linear? Yes.
GitHub? Yes. So that means the agent can have access to my authenticated account of said tool that I use.
The one mistake that you might get into is installing a bunch of stuff that you don't need. Like for example, let's see if I've installed anything I don't need. I pretty much installed everything that I needed, but like don't install Asana, You don't use Asana.
Don't install Dropbox. You probably don't use Dropbox or ClickUp. Right?
I have the presentation ones installed here. I barely use it. Don't be someone who just gets a new iPhone and you install all the apps and you never used them.
Use and download the plugins that are actually part of your workflow. So number one, plugins. Make sure the computer use one is installed.
Everything else is just based on your workflow. Number two, I wanna talk about the in app browser. It is actually fantastic.
Now, I don't know if you remember this, but OpenAI actually had a browser called Atlas. It was called ChadGBT Atlas, and it was the competitor to Perplexity's browser to Dia's browser. It was going to be the AI browser.
I remember the AI browser wars were at an all time high and OpenAI entered the gates only for them to revert back and nuke the app, but we find the browser now in Codex.
Now, this is your standalone browser, right? So I can go to youtube.com/@rossmike and I can make sure I hit the subscribe button a 100% and I can even see I'm logged in.
You know what? Why don't you hit the subscribe just like how I did? But number two is the session is tracked.
Right? So I can log into my Google account and it's there, but it's a lot more than that. I have this app running.
This is a simple to do list app built with SvelteKit and Convex. You see this cool annotation tool.
Now, sometimes you might let's say you wanna copy change or you want this font to be made bigger, you want this to be changed. There's a specific part that you wanna be changed. Prompting might be too much because specifying might be hard work.
Now you can use something like Whisper Flow to voice a dictate, but with Codex, it's as simple as clicking annotate. I can click here and I can voice a text, Hey, can you change the title to make sure to subscribe to Ross Mike because he's handsome? And then I can hit enter and here's what's cool.
I can keep stacking these. Right? I like the fact that it doesn't send this off right away.
So I can continue to stack annotations. I find that this buns a little too big.
Can we slim it down a little bit ? I hit enter right there. Okay.
So it's spelled Ozempic wrong, but it's whatever. So we can have as many annotations as you can see and I can hit enter. Now, this is going to be sent.
Now, on another thread, and we're gonna talk about threads later in the video. On another thread, I have this application called Pluto running and let's say I want my agent to test out this application. So this is what I'm going to do.
I want you to use the browser that you have. I have an application running in the browser. I want you to test agent creation and I want you to tell me if it works or if something can be improved based on your usage.
That was a terrible prompt. I literally forgot English for a second, but we're going to hit enter. And what I want us to see, hopefully, if this doesn't work, then I'm gonna show you how to guarantee it to work.
I want us to see the agent use the browser. As you could see, the cursor popped up in its own browser.
It's going to basically test out the app and it's going to let me know if my app is or so. Part of my workflow now is I will get codecs to run my application locally and to test it out.
So I just don't want code review, but I want user test review. This is basically the QA process automated. You could see it created an agent.
We're gonna see if the agent creation works. I like the name Codex creation test August 12. Coming back to the other thread, you could see that the annotation changes I requested have been complete.
And if I go back here, you can see that the agent is continuing to test this application. So the agent finished testing out my app and it worked. It worked and it even give me some valuable improvements that we can do and I could literally tell it to continue on building.
That's what we're here for. So so far we talked about plugins. Make sure to install the computer use plugin, and only install plugins that are part of your workflow.
In terms of the browser, the browser is great for annotations. Annotate the heck out of your app until it is the desired state and functionality that you want it, especially when it comes to UI UX. And another thing the browser is great for is telling the agent to test your app.
The third thing I wanna show you is computer use. The Codex agent is by far one of the best at computer use. But you know what also is the best?
Today's sponsor. Thanks to AI tools. My team and I are shipping like never before.
I mean, look at this. This is my contribution graph right here. Over 5,700 contributions in 2026.
And in the couple days, I've just been going ham. But it's not just me. It's my entire team.
We are shipping nonstop. It is PR galore. But I'll be honest, up until recently, waiting for our CICD pipeline to finish was the worst part of our day.
It just slowed us down so much. But would you believe me if I told you just this one line change saved us 40 to 70% time on our CICD pipeline?
And that's why Blacksmith is the best. It is the fastest way to run your GitHub actions. If I show you some of our workflows, you're seeing this a minute twenty eight, forty six seconds, fifty four seconds.
Some of these used to take three to seven minutes. Now they're all under a minute and thirty seconds. By the way, it's not just me who praises blacksmith speed, it's also Clerk and Mintlify.
In a world where speed is becoming the norm, thanks to AI, do not let your slow CICD pipeline be the blocker. Set up Blacksmith right now.
The link is in the description. Believe me, you'll thank me for it. So the best way I can explain to you how good the codecs computer use is, I'm gonna give it the most goofiest task because you're going to realize how good this thing is.
I want you to create a folder on my desktop, call it knock knock jokes, and in that folder, I want you to have three different folders, folder one, folder two, and folder three. And in one of the folders, randomly select yourself, I want you to have a markdown file with a knock knock joke.
Please don't pick a lame knock knock joke, let it be a hilarious one. So this is the most random goofiest prompt and I kinda wanted this to be the test for you to realize how powerful Codex's computer use is.
So I don't know if you could see, but you could see Codex's view over here, the agents view over here. Right now, it's using Finder and what's cool about the Codex, uh, computer use agent is I can continue to use my computer.
I'm not a blocker. I can continue to do my thing and it can use my laptop while I'm using my laptop.
Right? So like, if there's not a shared cursor. Right?
Like, I can move my cursor around. It has its own cursor. It's not going to be a problem.
And I could see here it opened the the text editor and it put a knock knock joke and the knock knock joke, I I didn't even get to read it. It's basically saving it. You can see the tabs here running and it's pretty fast too.
This is one of the fastest computer use agents out there. Alright. It says create a knock knock jokes on desktop folder folder one, two, and three.
Let's go to the folder. If I open the folder here, folder one empty, folder two empty. Oh, it's in folder two.
And then the joke is knock knock. Who's there? Control freak.
Control freak. Okay. Now say control freak who?
What the Alright. There's some things you can't trust AI at and that's clearly writing knock knock jokes. That being said, the computer use is powerful.
Now, this isn't necessarily just a developer use case. If anything, I'll be honest, I use the computer use stuff that Codex offers for like tasks, filling out forms.
I have a friend named Zavery, shout out Zavery if you're watching this video, my church bro, my brother in Christ. He works a sales job and he wanted to scrape leads of like he had a huge database of sites and he just wanted to scrape specific information or go on this site and get specific information. And I literally told him to use Codex and the Codex computer use agent literally went on every site and then created a spreadsheet, stored the information in a spreadsheet, maintained that spreadsheet, and by the time it was done, he had the plug in for Gmail sent him an email.
So this is more like knowledge work stuff in my opinion. Because when it comes to testing the app, the web apps, you're building a web app, I would use, you know, just the in a browser. But let's say you're building a desktop app.
Right? You can get it to test out a desktop app for you. So the computer use stuff is by far one of the best.
There isn't a harness that comes even close. Number four is thread maxing.
Now a lot of you wanted me to talk about thread maxing, and I want to make something clear. There are a lot of things in our space right now that sound cool, but aren't productive.
I don't want you to fall for the things that sound cool. I want you to follow and do the things that are productive. Now, when it comes to thread maxing, there's a particular mindset that I follow that makes thread maxing worth it.
Let me explain. Let's do the following. I want to build an agent, but I'm not sure which framework to follow.
There is the Vercel Eve agent framework. There's the Cloudflare agent SDK, and there's Anthropix and both OpenAI's SDK, and I think Efct also has an agent SDK.
I don't know which one to use, I don't know what's the best. I want you to do the following. I want you to review the docs, I want you to build a simple agent using every single framework, but I want you to do this in separate threads.
Each framework gets its own thread, but I want you to present research and facts as to why this framework's the best. Once you're done, let me know which framework we should go for And for further context, this agent is for a construction company that's looking to build an agent for their operations.
Lately been getting a lot of construction companies reaching out to the consultancy regarding agents. So just, um, little context to share with you why I pick construction. So check this out.
It's running. Look at the threads it's spinning up. It's spit up one for effect, one for the OpenAI agent, for the Anthropic agent, for Cloudflare, and for Vercel.
So like if there's like this type of research that I need to do, I'm spinning up threads. So thread maxing, and I'm gonna share with you when I thread max. First of all, our research on competing things like frameworks, you know, infra, whatever it is.
Number two is, let's say there's two ways of building something. Right?
There's let's say there's two architectural decisions or two design decisions you have to make. I will literally tell the agent, I want you to spin up two separate threads and I want each thread to build it, use a work tree and then I want you to have a running version on the browser for me to review. Right?
So if there's like I'm on a fork on a road, I don't know which side to pick, I will literally tell the thread I'm on to spin up through two threads to build it and then I can review and then we could stick to the one way of building things. Another side thing that I wanna bring out sort of unrelated to the flow of things, but that Codex is really great at.
Codex is really great at configuring just your computer. Is there a certain thing not installed? Are like you wanna check how much memory is being wasted?
You're setting up a home lab and you want certain things set up. Codex is fantastic at all these type of things like so right here, I have Ghostie running Herder. This is actually my Linux box, and I'm using the Codex CLI to just set up all the network stuff, to build dashboards for me to check things like codecs is amazing.
Obviously, we're talking about the GUI here, but even the CLI is really, really good. But in particular, when it comes to like using the machine that it's on, I can't recommend this GUI enough. Now, while the threads are cooking, why don't we talk about model selection?
What model do I use? Right? I'll be honest with you, I'm using SOL for everything and that's because the subsidies are great.
I'm on the $200 a month plan. If you are not and maybe you just can't be using SOL all the time, Tera is good, Luna is good. You just have to think of the type of task you're using it for.
I'm generally building an application, diagnosing something on my network, doing something that requires a little bit more brain, so I'm living in the SOOL department. Now, effort, effort, effort, effort. For effort, I'm using medium and extra high.
And there have been times, ladies and gents, I've used ultra. Now there was a little bit of an issue with ultra a couple weeks ago, burning tokens and people complaining. I used it a couple times and it was a okay.
It was pretty good. It's not no cloud to cloud work flows, but it's really good. But in all honesty, I think you can live a pretty happy life if you know the Lord Jesus.
Hallelujah. And you live between medium and extra high. I'm always switching between medium and extra high.
I find that this is a happy medium for me. If I find the task to be something on the simpler end, medium. If it requires a bit of brains, extra high, that's where I live.
GBT 5.6 Soul. I have speed fast because I'm entertaining you lovely gentlemen and the three ladies including my mother and wife who watch this. But to save tokens, you can go on standard.
So the thread is finally done. Right? After this prompt and after all these threads got to work, I got a final response that says I recommend Cloudflare Agents as to here for the construction operation agent, it's a durable object plus, it continues on, it ranks it in this order, I have a file to review.
Like this is just great for there's a fork on the road, you have to make a decision, there's a bunch of things that you have to research, you tell the thread to spin up a bunch of threads, come back with the result, . And we talked about model selection, g b d 5.6 sole medium and extra high. Now, number six is slash goal.
Now, I didn't like slash goal in the beginning because people were using slash goal to build full on projects, but that's not why you use slash goal. Slash goal is used in the following. When you have a defined goal, right, and it takes multiple steps to get there.
Most recent way that I use this, for example, is I had a website, a client site, and the client site was slow, and we want it to be super fast, blazingly fast, click and it's there. So what I did was slash go, mind you, Codex has a browser.
So I prompted saying, use the browser. I can tag the browser and run a Lighthouse score test.
Right? Lighthouse score basically determines how accessible your site is, how fast your site is, and the agent can easily do that. And I said, make this site blazingly fast and get all hundreds.
And then I hit enter, and slash goal was on, and it worked for, like, three hours. The result was a site that had a full lighthouse score of a hundred hundred hundred, and that was blazingly fast. Now slash goal only worked because I had given it the final goal, right, like what the finish line looks like, but I also gave it a feedback loop.
The feedback loop was the lighthouse score. The score had to be hundred hundred hundred because there's a way for it to check its work. There was a feedback mechanism, and there's a defined goal.
My slash goal run wasn't garbage. Most people, and I'm sure you've experienced this. Your slash goal run is garbage, and the reason why it's garbage is because it doesn't have a feedback loop.
So in order for your slash goal run to be successful and effective, you need a feedback loop, you need a defined goal.
Right? Combine these two and you're gonna love gold. I wouldn't use it to build applications unless it was some random internal tool.
I used this to build Mvrun, which is basically my way of syncing environment variables to my one password, I have one password set up on all my machines, and that way I don't have to copy paste environment variables. Right? I started this with slash goal, and then I tweaked it a little bit.
Something like this internally, sure. I wouldn't use it to build apps, but I would use it to improve certain parts of your apps that you can have a feedback loop attached.
Therefore, it's not just generating slop. Goal is one of my favorite tools. I use it only if I have feedback loop.
And finally, I've I keep mentioning this skill, the grep loop skill. This is the skill by Greptile. Again, Greptile do sponsor videos.
This is not a sponsored video by them. But the reason why I share this skill is because this skill sort of works with like gold.
And basically, what the grep loop skill does is it improves the PR until grep tile gives a five out of five confidence score with zero unresolved commits. Triggers GREPTILE review, fixes all actual comments, pushes and reshelves, re triggers review and repeats.
Use this when user wants a fully optimized PR against GREPTILE's code review standards. So this is basically like a slash goal. The goal is I want a GREPTILE five out five.
I want you to continue to fix whatever gremtal gives you, and until you get a five out of five, don't stop. This works because there's a feedback loop. And this is why I love gremtal because gremtal makes it easy for me to optimize review, for me to make sure my PRs are as less sloppy as limited slop.
It's like a . It it just it's one of the best things out there and I want you to think of slash goal that way. The feedback loop is important.
Leave in the comments down below feedback loop. And the one final thing that's cool about goal is while it's running, you can do other things like read your bible. You haven't been reading.
You should read. Why don't I read a verse for you? Let's just pull up a random verse.
Actually, is perfect. Proverbs 27. Proverbs 27 verse one says, don't brag about tomorrow since you don't know what the day will bring.
Let someone else praise you, not your own mouth, a stranger, not your own lips. You know, sometimes we have a tendency to brag. The Bible says the humble will be exalted.
Now, we talked about plugins, browser, computer use, thread maxing, model selection, and number six was goal.
And number seven that I wanted to bring up was actually t three code. And the reason why I bring up t three code is you can use your codec subscription within t three code.
And there's a couple cool things that t three code has in particular, if you have different machines being able to work on your different machines is excellent, but really the most important and impressive part is the settle inbox version. Like, I'll show you an example.
This is better than this, especially when you're working on multiple projects, multiple threads.
But if you want me to do a separate video on t three code, let me know in the comments down below. Ladies and gents, that's pretty much it. I hope you enjoyed this video.
I hope you learned how I use codecs, and I want you to go maximize the heck out of it, especially while the subsidies last. But thank you so much for watching. I hope you enjoyed this video.
Make sure to like, comment, subscribe. Make sure to subscribe because we're almost at a 100 k. I'm gonna do something special for a 100 k.
Thank you so much again. I'll see you in the next one. Peace.
The Hook
The bait, then the rug-pull.
Ras Mic opens with a direct callout: if you're using Codex as a prompt box, you're missing out. What follows is a run of seven concrete habits, from plugin discipline to a review-until-passing skill, that turn a chat window into an agent that browses, tests, files, and configures on its own.
Named ideas worth stealing.
00:44list
The Seven Codex Power-User Habits
- Plugins matched to your workflow
- In-app browser for annotation and QA
- Computer use for desktop/knowledge work
- Thread maxing for parallel comparison
- Model and effort selection
- Slash goal with a defined feedback loop
- T3 Code for cross-device access
The video's own numbered structure, seven stacked habits that separate a power user from someone who only ever prompts once and waits.
Steal forany agentic coding tool, not just Codex, as a checklist for moving from single-shot prompting to a real workflow
16:53concept
Goal Plus Feedback Loop
An unsupervised, multi-hour agent task only stays on track if it has both a clearly defined finish line and an objective, checkable signal of progress; without the feedback loop, the same mode produces low-quality output.
Steal forany long-running autonomous agent task, deciding when to trust it to run unattended
18:06model
Grep Loop Skill (Greptile)
- Triggers a Greptile review
- Fixes all flagged comments
- Pushes and re-shelves
- Re-triggers review
- Repeats until a 5/5 confidence score with zero unresolved comments
A named skill that applies the goal-plus-feedback-loop pattern specifically to code review, using the review score itself as the feedback signal.
Steal forkeeping PR quality consistent without manually re-requesting review after every fix
How they asked for the click.
VERBAL ASK
19:36subscribe
“Make sure to like, comment, subscribe. Make sure to subscribe because we're almost at 100k.”
Casual, running-bit ask (also planted earlier in the intro) rather than a hard mid-roll pitch; paired with a promise of "something special for 100k."
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
