How I run Codex from my phone and 30 scheduled automations

Hey guys, hope you're well. Today we're going to go over how to maximize the value from your AI coding agent slash agents. I think this will be quite a long video.

It's basically going to be all of the information and workflow improvements I've made over the past couple of months. Topics that we're going to cover is first off what the actual problem is. Then we're going to go over a mobile application that I've made that actually allows you to communicate with Codex whilst you're out and about using a direct OpenAI product.

Then how to maintain a very clean codebase that's AI agent friendly to avoid wasting tokens. Test-driven development. Codex automations I think is one of the most interesting things I've not actually heard anyone talking about. skill level memory, gathering feedback to improve your agent automatically, and then just some other notes and things that I've been been going over. So, first off, to give some premise and introduction, my feeling at the moment is that everyone is completely obsessed on the completely wrong stuff.

The questions are always around, oh, what about this versus this? What makes this better? And it's because of these YouTubers posting these videos every day saying this changes everything, making hype around something that really makes no difference at all.

My opinion with just my experience, the only tools that ever really change everything are the releases that come natively. Companies like Anthropic or OpenAI. Whether that's a new model or they release a new tool, they are generally the gamechanging tools like Hermes or Claude. like I I do I use them both but it definitely doesn't come under gamechanging feature or anything like that.

So don't worry about what wrapper are you using or what agentic harness are you using. It really comes down fundamentals of having good prompts and that also means building an environment where the AI agent is actually able to understand what it is that you're trying to accomplish and how it should implement that. So you need a clean structure.

It doesn't need to be over complicated. The whole point is about making a really good environment for your agents to be able to work in a healthy codebase, good structure, good prompting, testing, automated checks when you're deploying your code and just well documented and structured codebase. The other thing I haven't added in here that I should have is I also think that there's just so much wisdom and knowledge in the tried and tested format of a traditional software development team where you break up the different roles and you would have a project manager who will keep things pushed through the pipeline and has a broad overview of the whole whole project.

You have your developers. You would have a solution architect who will architect the solution of a project as a whole of how is this going to be built and then a technical architect who would do something like reviewing a spec and reviewing the actual implementation plan to make sure that it follows rules and structure that you've defined in your codebase before handing it over to a dev agent. So I think that extreme ownership really applies quite well here.

People are saying things like this model is shit I would generally think it's actually a human issue where either the issue is in your prompting or your environment. So how do we actually solve that problem? And then basically the point is don't get sidetracked by this changes everything and by what all of these YouTubers are doing. follow really healthy, clean principles of maintaining a good codebase and a good environment with clear rules, clear documentation that doesn't blow your agents context window, which is one of the main things you're going to be fighting against.

Now, the way that I like to do this is I have automated checks of a file size. So, the number of lines in any file in my codebase except for tests to be 600 lines. This can be done in a completely automated way.

So if any of you are starting new projects, I encourage you to enforce these rules before you start writing code because otherwise you'll kind of go through what I went through where you have a two week refactor on your hands. You're just blowing through tokens refactoring everything. There's a massive payoff at the end of the day.

The reason that it works so well is because you haven't got these limitations and guard rails in place. What ends up happening is you get these god files. I had some files that were 5,000 lines of code.

And what happens is your agent is searching for something, opens that file, just blows up the context with unnecessary crap that it doesn't need. The outputs are also so much worse because it's so hard for it to keep all of the context in memory of everything that's going on. This one huge file that you have.

And I had loads of these files all over my codebase. By reducing file size and breaking up into smaller into a leaner codebase just get you get much higher accuracy. You use a lot fewer tokens.

It becomes much much easier for the agent to be able to find what it needs. If the codebase is structured in a logical way. So 600 lines max file size 80 lines max function length and the formatter that I'm using at the moment is Ruff and also Pyright.

I'm able to get so much more work out of these agents because the codebase is structured in a much cleaner way and they're able to find exactly what they need without blowing through load of tokens basically. Now we're going to get into the interesting stuff. So I've built out this app called Codex server and this is using Tailscale to communicate with my machine here.

OpenAI actually offers something which is a server that you can just run on your machine that exposes access to all of your conversations. You can communicate back and forth to the conversation. So from my phone here, I don't have to be on the same internet.

I can be out and about. I can be on a beach. I can be in bed able to communicate with these agents.

And it's very very fast. It's like instant thinking he's giving me a response. So this front end was made by by GPT in one prompt.

So I've been looking for a solution like this for a while because basically I've been feeling this pressure that you need to sit by your laptop the whole time with 20 windows open because the leverage that I'm getting because the amount of value that I'm getting is just so insane. I've just felt that I needed to be glued to a screen with all of my different agents open in order to be feeling productive. So, I've spent the past three months just glued to a screen, 16 hours per day, waking up, going straight back in, doing it all again.

And I I'm happy. I'm to be honest, this is one of the happiest points of my life cuz I've just been able to build so much stuff. So, I'm really happy to do that.

But still, I want the flexibility of if I just go out that I'm able to respond to an agent to keep things moving forward. Now, I also do a lot of walking as well. I like to walk because I'm able to think.

So, the next thing that I've built out is I've built out a two-way calling. So, this is built on ElevenLabs. So, you have transcription through here.

Let's check if it actually works. I do it from here. I can transcribe a voice note and it will now appear here.

And it transcribes in real time. You will see the text is coming in there in real time because ElevenLabs have a real time text to speech. So, I'll send that.

But then what I'll also do is a call which I'll turn on now. You speak transcription in responsible. Anyway, you get the idea. but it reads out any messages that comes in and in real time I can be walking with my AirPods in and I'm able to speak to it and just have a conversation with all of these AI agents that are operating main machine.

Another thing that you can do as well so here I've got it called Codex mobile. So you can have a high-level orchestration agent. I can say, can you you go through all of my existing chats and look for common points, common complaints I have that we could turn into docs or skills.

So, this has access to all of my conversations that I've had in Codex and because it's exposed through the server, we'll just go through read all of the chats that are open here and it can just go and extract. It will do pattern mining to basically look for improvements it can make based on common things that I complain about. But you can also have it message the agent.

So I I find it's better than than a sub-agent structure. I've only built this out about 2 days ago now. So it's still quite new.

I'm still playing around with it. But the mobile app I absolutely love aside from my terrible demonstration here. Essentially you just have real-time communication from your phone.

It's synced automatically. Everything is handled by OpenAI just running a server basically locally on your machine exposing it through Tailscale. I can also put the prompt in the description that I used for being able to set this up just compared to all of the other tools that are available for doing this.

One called happy. There's a Y Combinator backed tool called Omnara. There's another one called Remodex.

And I've used all of them because I've been trying to solve this problem for a long time. and you have a few issues using other tools. One, you have a privacy issue in my opinion. I prefer as few people to have access to the codebase possible and what my agents are doing.

So by working natively with OpenAI, of course, OpenAI know what you're doing, but know that anyway. You just limit the number of parties involved. So that is in my opinion a really big security benefit particularly if they've got vulnerabilities their side or whatever. just increasing your exposure by using these third party tools.

But the other point is they're just not that stable compared to this. I I cannot believe that like stability and speed that this is able to work at. Any of you that are using Codex a lot, I would encourage you to basically build a mobile app companion.

This is just built as a progressive web app. So, it's built in the most simple way possible. Essentially, it's just saved as a local website, my device from Safari. that I can still do things like sending through push notifications.

And right now it's an MVP, but eventually I will actually turn this into a native iOS app, not to be published on the App Store, but you just get a lot more benefits in terms of things like background refreshing. Notifications become smoother. The calling logic also becomes a lot easier to do as well because at the moment with the way that calls are done, I need to have the phone open in order for the call to stay there because it doesn't support background audio when the phone locks.

Whereas if you're doing through an actual iOS app, that behavior all becomes much easier to implement. But considering it was done with one prompt and then for the past two days I've just been doing minor tweaks and integrating ElevenLabs and sorting out all of that streaming and a lot of intricacies in behavior of allowing me to interrupt it if it's reading and so on. But I'm really quite happy with where it is.

Next up, this is one of the most interesting things that I just haven't heard anyone talk about that I'm going to go into now, which is these Codex automations. So you can set it up here and I can just say in automation of whatever go and find a bug in the codebase or you can actually set them up. You can have an agent set them up for you and exist in I think it's automations.toml .toml and yeah it's going to read through but can have it set up new ones or make improvements to them or turn other ones off or change the scheduling or make a different prompt.

But essentially it's a prompt that runs on a schedule. So I can choose to run it on the main code that I have locally. Make a worktree.

What project do I want to run this on? When do I want to run it? And what model do I want?

High do I want the reasoning to be? Share all of the ones I have here. So some of the more interesting ones would be looking for gaps in testing behavior.

So every day at 3:00 p.m. this is going to run. It's going to scan the codebase and look over the critical paths in particular like the back testing paths. Going to basically make one change and you can see the different runs going on here.

And it doesn't deploy any code to production. It just opens a pull request and I can either have another agent go and look at that pull request and check does it like the change? Is it a valid change?

Just be disregarded if I don't like it. So for me, these tokens are so cheap to run. And if every day this runs and makes a 1% improvement with no input from me at all, every single day, it improves our testing behavior by 1%.

Again, because of the type of project that I'm working on, testing is just so so important. Harden silent failures. So this looks for specific paths that haven't got good logging or could be failing without us realizing.

Scans the entire codebase every day. It can use sub agents. It scour through everything and it looks for what it believes the highest leverage thing to improve is.

Makes that improvement, opens a draft PR to be reviewed by myself, by an engineering manager agent. Code quality hardening. One I really like is remove dead code.

So it looks for paths that have been superseded, that are no longer relevant, that aren't being used anymore, that are just completely unnecessary. It could even be scripts that were put into production before that really needed, that were one-time behavior. one I really like which is skill improvement. I've got a few on skill improvement.

It goes through all of my GitHub PRs. It can review the comments of everything that has been merged. It goes through Linear issues and it will look through the comment section of what agents were talking about and what problems they were having and it can look over let's say from the past week.

It understands what were commonly occurring issues that agents running into and it will go and turn those into either documentation or it will create a new skill or update our skills in order to just make sure that future agents aren't wasting time on the same problems. It works well on things like setting up of the environment. Python you have to use .venv virtual environment on a Mac.

And there can be different issues with different versions of Python depending on what I'm running on some of the machine learning tools. It only works with specific versions of Python. It's just about it automatically understands that based on what agents were complaining about or what they said, what issues they were having in the comments.

It pulls all of that information and turns it into a skill so that future agents don't have that problem. And therefore you're not going to waste tokens on agents continuously running into the same problems walls because you have a constantly improving agentic codebase. Bug hunter that is actually another really good one.

The thing that I do is I have one monitor my Better Stack and Sentry errors. It runs quite frequently. I think it's every 3 hours and it basically checks are there any new errors in Sentry or anything in a Better Stack logging that's running in production that looks like it needs to be fixed.

Complexity reducer I like a lot as well reducing the size of the codebase by looking for overengineered or code paths that can be simplified. Doc synchronizer scan the last seven merged PRs and make sure that all of our documentation is still up to date and is still relevant. Documentation validator, another great one.

I'm full of good ideas. Validator goes through our documentation, follows it word for word, and makes sure that it's able to basically get something running based on what is specified in our documentation to prove that our documentation is actually up to date and working. This is all automated.

So these are codebase hygiene things or just maintaining a generally healthy environment, keeping the documentation up to date, keeping CLAUDE.md up to date, all these different things. it can now just completely be outsourced to Codex automations or Claude do also have their own equivalent of automations as well. So you can also run skills from there as well. So the next thing that I found to be really valuable is to have a skill level memory.

This is basically notes that persist between runs. So you have all of your documentation, but then you don't want to put everything into documentation because then the documentation ends up becoming bloated itself. So let's say that I have the documentation validator that can have its own notes section that exists within that skill and it's a hidden folder.

So it's not actually part of the codebase. And whenever the docs validator runs, it goes and checks what have all the previous agents ran before. So it's not doing duplicate work.

It knows on this day at this time that path was already checked and validated. Same with something like a bug checker or complexity reducer. So there's a ledger of the history of runs and the previous scope of work and any notes from the agent.

So I have about 30 skills or something. I've got about 30 35 automations that are running that each have their own knowledge specific thing. So you only use this if the knowledge is only specific for that specific skill.

Claude's done us a nice nice diagram here. So agent boots, load the notes, read the notes, run the task, update the notes at the end. This is just on making our skills smarter.

So our agent runs, it goes through and scans what are common issues that previous agent ran and then that will automatically picked up and put into the skill so that the future agents don't have that same problem. It's about having a self-evolving, self-improving system. But this is all happening without me needing to do anything, right?

Completely autonomous. How it actually improves. One other thing that I found so interesting is these agents are so capable of performing so many different tasks and it really comes down to you as the orchestrator of these agents, as the manager of these agents.

How good are you at extracting value? Because they're smart and they can work on all different sizes of problems. The reason that this became so apparent to me is I've been working in a co-working space recently.

And when you're like walking around, you'll see what everyone has open. In some way or another, every single person is using AI. I don't think there's a single person in that office that doesn't have either GPT or some coding agent open.

A lot of people designing websites, a lot of people doing copywriting and writing articles, which I would consider to be like much kind of low lower value tasks. programming it is on the higher end of that but then it comes down to how many agents are you actually able to manage concurrently but you can like for example I'm working on a lot of machine learning I haven't got a background in machine learning I know nothing about machine learning but because of these agents I'm now able to build out these machine learning data engineering pipelines for running back tests for an agentic hedge fund previously to build out something like this you would have needed mathematics, CS, PhDs, data scientists, data engineers. The cost of quantitative developer in New York is going to be, I don't know, half a million dollars per year.

And it's not to say that they're even on the same level as that. Maybe that is debatable. But are they 80% there?

I would say so, or at least they're kind of not far off. Particularly these models like GPT-5.5 Pro. The thing that I find so interesting is we're all paying the same amount of money for these tokens.

And really the thing it comes down to is how big of a task, how valuable of a task you able to actually get agents work on. Anyway, next up going on to test-driven development. So this has been a new implementation, the new skill set that I've been using.

I really like this guy's um skills by the way as well. I really like his brainstorming workflow. Um, so I've been using this quite a lot recently and it has test-driven development built into it.

It does is you give a prompt or you go through the brainstorming workflow with it. It understands what it is that you actually want to build. It writes tests to validate the behavior before it actually writes any code.

Before it starts the implementation, it writes tests. The tests are going to fail obviously because we haven't actually written any code yet. Then it writes the code, gets the test to pass to validate that the behavior is working as expected and that is how it's able to prove that the implementation has been done and it solves this exact problem here.

A few other benefits of it as well is depending if you're doing unit tests, it verifies that that code is working in future as well. So you have the guard rail just to validate that everything is still running as expected. But then I've also been doing a lot of stuff because I'm doing back tests again for an agentic hedge fund.

You need to make sure that your changes haven't impacted previous behavior in an unexpected way. So I'm doing a lot of regression testing and giving fixed data cache or dummy data and then it will run the code and it should make sure that the outputs are always the same. So if you're feeding in the same data, it should always be the same output and if it isn't, then you know that something has changed from your codebase unexpectedly and my checks won't allow it to be deployed production.

So yeah, they are main things I've been going over. This mobile app I strongly encourage you to use and then just working on setting up system that is automatically improving so that your agents don't keep running the same things. Another really interesting concept that I've coming across recently is I've just been working co-working space.

You just naturally look around and see what other people are doing. You're spending all day there. Just looking around at what people are doing.

Just thinking your job so easily automated and it's not a case of they're not using AI and they don't know about these tools. They are using AI. they just don't have an understanding of bigger picture of how to really get value out of the AI, how to really automate entire processes exactly like what I've been talking about for the last 30 minutes. And so one perfect example of this is there's a guy in the co-working space, young guy, he must be there for eight hours a day with Notion open on one screen and then on the other screen he's using Kling and video editor.

In Notion, he's copying a script, then also a prompt, pasting it into Kling, dragging that Notion task as in progress, waiting for it to render for a couple of minutes, downloading the output, uploading it to Google Drive, uploading it to done, and then I think he just in Google Drive. So, he's doing it for a supplement thing. And he's got these old men that are really ripped, just shredded with loads of hot young women around them trying to sell these diet pills, whatever.

I just look at this thinking this is a workflow that could have it automated a couple of hours. It's they have no idea of this. Like I'm sure in many ways he thinks it's very clever because he's using Kling.

He's making these things. I think he works as an affiliate or something for these supplement things. So he thinks very clever because he's utilizing AI but he is still the human in the loop.

The limiting factor becomes him only making one video at a time. He's sitting there for 3 minutes at a time waiting for it to render. And with a really easy script, we could scale him out to have a hundred of him rendering these videos all simultaneously and essentially would be completely replaced.

And this is just one example. There's loads of copywriters as well that feel the exact same thing about. And this isn't to scare people or to look that there's an opportunity here.

This is here already. It's not a case of this is coming 6 months down the line. We're far beyond that point.

This technology is already here and it is happening. So for those of you looking for new opportunities of ways to make money, I would even encourage you to go to a co-working space for a day just to walk around and just to see what everyone's doing because you'll notice really clear trends of in the co-working space I'm in, everyone is either designing websites in some way or another, designing websites with AI, some of them doing in different ways, they're developing, but generally just with one terminal at a time.

There's one guy I've seen at most about six terminals open. So he has got more concurrent workflow running. But that's not to say that not about output of code about volume.

It's really about quality, right? But you can look for these opportunities of what are these people working on and they're obviously making money from it some way or another if they're getting paid a salary or getting paid in commission and you can look for these things. You could just automate easily a day could have this automated.

Another concept that I just like at the moment as well is to throw compute at your problems and just see what it comes back with. Any issue that I have now and it doesn't necessarily mean compute as in coding agents love GPT-5.4 Pro and GPT-5.5 Pro it is now are incredible models and I throw everything at them. If there's a big thing, not even if it's something I'm stuck on, just something like an idea that I have, something that I'm thinking about and you just set it off in research mode and it goes off for about 45 minutes.

It is incredible. Really, really great models. thing that I find it is best at is having knowledge of what tools are available. It scraped the whole of Google and absorbed all of this information.

If you say that you're having a particular problem and what is a good approach to go about solving it, it doesn't need to be technical. It can be absolutely anything in your life. Play around with it.

I probably send about three to four prompts to it per day, bearing in mind it takes 45 minutes to reply. So only if it's a big picture idea, make a voice note and send it to it and let it run off and just see what it comes back with. For example, the mobile app that I showed here, I was basically telling GPT-5.5 that I was trying out all these different tools and I just wasn't able to find anything good and it just told me, do you know that OpenAI has a native server that you can just expose, build it on Tailscale.

It gave me a prompt for it. In one prompt, I had something working. So it's mind-blowing. real skill I think to focus on is not down to engineering or how good are you at prompting.

How good are you at getting value out of these agents? How good of an orchestrator are you? That is the only thing that I think we should all be focused on.

It's exactly what I'm focused on now. And by orchestration, I mean these different things like Codex automations. These things are just running for me automatically whilst I sleep at 3:00 a.m. at 6:00 a.m.

So, I just wake up to this inbox of tasks that have been completed overnight. I sign off on things that I'm happy to be merged. I deny the other things I'm not happy with.

But basically building an environment where the agents are able to work in that way. They're able to work autonomously, come up with new ideas without breaking anything because they all work in worktrees. So, they're in isolated environments.

And that is where I think that you start to get into the $1,000 per hour and $10,000 per hour tasks. But another way to even look at it is like if someone's doing copywriting with AI, okay, it's definitely a step up from doing it manually yourself. But depending on the level of copywriting, what is it really per hour task?

Probably less, maybe more for a very high level copywriter. So it's better to use AI that than not to use it. Do you want to throw your tokens at?

You want to find the things to push your tokens at that really incredibly valuable looking at things like paying jobs. What are the highest paying engineering jobs and how can I utilize that in my codebase because these agents are fully capable of doing it. It's just about you having the the imagination to the most out of them.

The other point as well is that I still find that the limiting factor is people. Basically, it comes down to your knowledge. When I was talking about this guy here, the reason that he's doing this all manually is just because he doesn't know any better, right?

Hasn't had the experience of building out a system or even understanding that it's possible. And this is a really simple example. So, we all have access to the same tools, to the same level of intelligence on the general level.

Some of us are on higher plans than others, but as a general rule, we all have access to the same level of intelligence. So really what you need to be focused on is how can you maximize the amount of value that you're able to extract from these agents and that falls back to you. It falls back to what your capabilities are in terms of what you actually know.

It's the same with developing particularly with APIs. An API isn't hard to integrate. What it really comes down to is what APIs do you have knowledge of?

So what do you know is possible to do with APIs? If someone knows about an Instagram automation API and work in OF or in marketing or whatever, they're in a much better position than someone doesn't know about it. So simple example, but basically you want there's still huge amounts of value in your own personal knowledge in my own personal knowledge from building a team before out to staff. the lessons that I learned there through a lot of failures in management in terms of people where I I really learned a lot about wanting to maintain a lean organization because we I don't really like managing people.

Then all of my experience in in marketing, in automation, in systems building, in APIs, in all of this different stuff is now just more relevant than ever. And I just find it so so useful. It's not until I was looking around the office realizing that we're all using the same tools.

But in terms of the level of value that we're extracting per token, we're on completely different levels. One other really interesting thing because people always ask why I don't like Claude Code. I do like it.

I like it for front-end stuff, but just for backend stuff, it's just nowhere near as good. And when I made that video about 2 months ago, people were saying like, "What are you talking about?" And then there was like a couple of people who agreed with me. And I think the reason for that was because YouTube AI coding universe was everyone was talking about how good Claude Code is and no one was talking about Codex.

And now I would say that it's completely shifted and everyone now agrees. And what that tells me is that the these YouTubers aren't really building. They're building for their YouTube.

They'll use these tools on a very small scale enough to make a video. They don't have the depth of understanding of actually using these in real world scenarios. The people who are actually on the ground in the trenches.

All I spend my time doing is testing everything, testing new models, testing tools as they come out. But the thing that I realize is it's best to just stick with the real foundational stuff. But I just love tinkering with things.

It's just my hobby more than anything. But it was just so apparent, blatantly apparent that 5.2 Pro on anything aside from front-end development was a better programming model was anyone who was saying differently at that time I would just say that hadn't done a direct comparison anyway this is one thing I do really like in Claude just to show you that I don't complain about it all all the time right Claude Teams so what we can do here is can you set up a six-man research team to look for the best back tests you are the team leader use team members, not sub agents.

Now, more than anything, I just like this because it's quite cool. You can sit in your co-working space, and you're going to have a load of different windows open automatically. So, I'm not going to touch anything.

This is using cmux, which is a fork of Ghostty, and basically, it's going to launch a load of different terminal windows here, and it's going to need a minute to set up. One other thing that I do really like at the moment is computer use. So I can say go to Google and search for latest news.

So we'll run two tasks in parallel here whilst this thing is getting set up on the back end. Use Claude at the moment a lot for research because I like this Claude Teams feature because all of the different agents are able to communicate with each other and you can broadcast messages out. This is Codex computer use.

So this is going to open Safari for me and it's just going to do a search latest news. Very very simple example. I had it order a coffee for me the other day to my exact table and it was on not the cleanest interface let's say and it did a really really great job.

So I really like computer use here. You can see that this is running launching sub-agents. And the thing that I like about cmux is you can visualize all of the different sub-agents because they open all within different terminals.

So when you run something like this in a co-working space, you get quite a lot of people looking over thinking like what the what is he what's he doing? Because you can launch I think I've had eight or 10 of these running at once. It's just going to launch them all.

And this on the left hand side is the orchestration agent. So he's just going to oversee everything. And I can also have him send a broadcast out to all of the agents.

He can communicate with each of them individually, tell them to work faster or whatever it is. If you want to change something in the workflow, he can broadcast something out. All of the agents will then receive that message.

All of these agents, by the way, are able to communicate with each other as well. So you can have a shared information that happens in real time. So if I have all of these doing research, they can report into each other their findings to avoid things like duplicate work.

And now I think he'll send out a broadcast to all of the agents. Anyway, yeah, in terms of workflow stuff, have been using Ghostty at the moment as well, which I which I really like. Text rendering very very clean.

Yeah, very big fan. So I've moved to that away from from Warp. The reason for that is I just found Warp to if you're running tons of agentic terminals.

Warp becomes very very heavy on resources on RAM and Ghostty is just a lot more lightweight. But now because of this phone app I've just been doing a ton of work off of my phone. So yeah, I think I'll leave it there.

I'm sorry for the format of this video as well. I feel like it was all over the place. I do feel that there's a ton of knowledge and information in there.

Most of this stuff I haven't seen anyone talk about before because I just don't really see many other YouTubers that I look at that I think they're really in the trenches in the arena doing stuff. They're kind of monitoring from the sideline and just looking for the next clickbait video to put out there with no real deep insights into actually like the practices of engineering or the practices of these workflows. Any questions just let me know.

I'll probably open source some of this stuff down in the description as