23:07Cole Medin · Tutorial
Watch This If Your Coding Agent Is Ignoring Your Rules (You Need Hooks)
A walkthrough of why CLAUDE.md rules get silently skipped, and how to turn the ones that name an event into deterministic hooks instead.
August 27thThe argument in one line.
As coding agents, the models underneath them, and the workflows wrapped around them all improve in parallel, autonomous 'software factories' that turn a planning document straight into shipped code are becoming realistic for real teams, not just a hypothetical.
Read if. Skip if.
READ IF YOU ARE…
- A developer or team lead curious whether AI can realistically run planning, coding, review, and deployment with nobody reading the diff.
- Someone already comfortable with 'AI writes most of the code, I review everything' who wants to see what the next level of autonomy looks like.
- A builder interested in trying an early-alpha open source system that installs itself from a single prompt pasted into a coding agent.
SKIP IF…
- You want a finished, polished tool to run in production today — this is an alpha the creator is still actively stress-testing.
- You need step-by-step setup instructions — this video is a progress update and an invitation to build alongside him, not a tutorial.
The full version, fast.
An AI software factory takes in a planning document and ships working code out the other end: agents handle planning, building, review, merging, and deployment with nobody reading the diff. The video walks through the five levels of AI coding autonomy, from manual coding up to this fully hands-off 'dark factory' stage, then covers a year-long experiment that shipped a live AI tutor app built entirely by an autonomous pipeline. The argument is that this is realistic now because coding agents, models, and workflows are all improving in parallel, not a hypothetical for the future. An open source, install-by-prompt version of the factory is being released for anyone to try, though it's an early alpha built one planning document at a time.
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 – 02:37
01 · A PRD Goes In, Shipped Code Out
Defines the AI software factory concept: a planning document enters a pipeline of agents that plan, build, review, merge, and deploy, with no human reading the code before it ships.
02:37 – 03:52
02 · Dan Shapiro's Five Levels of AI Coding
Walks through the driving-automation analogy for AI coding autonomy, from Level 0 (AI as reference tool) through Level 3 (AI writes most code, developer reviews everything) up to Level 5, the fully autonomous dark software factory.
03:52 – 05:26
03 · The Experiment: My Dark Factory and Dynachat
Recaps a year-long experiment building a full dark factory that produced Dynachat, a live AI tutor app grounded in the creator's YouTube and course content, without a human ever reading its code.
05:26 – 07:13
04 · Sponsor: DataImpulse
Sponsor read for DataImpulse, a pay-as-you-go residential proxy service, demoed against a product-scraping script that gets rate-limited on a single IP and runs cleanly once routed through the proxy.
07:13 – 08:45
05 · Why I'm Changing How I Teach
A candid admission: past content always taught viewers to build tools from scratch, but none of it went viral the way fully-packaged open source projects do, so this project will ship as something people can install and run immediately.
08:45 – 09:46
06 · Inside the AI Software Factory
Introduces the open source AI Software Factory repo, built on top of the Archon workflow engine, installed by pasting a setup prompt into a coding agent.
09:46 – 11:37
07 · It's an Alpha, Come Build It With Me
Frames the release as an early alpha, not a tutorial-ready product, and invites viewers to follow along as the system gets built out and refined over time.
11:37 – 12:54
08 · The Projects I Need to Build With It
Explains the need for real test applications alongside the factory itself, revisiting Dynachat as the first proof point.
12:54 – 14:08
09 · Why Video Games, What's Next
Explains the choice to stress-test the factory on video games instead of web apps because games keep offering room for more complexity, then previews upcoming channel content.
Lines worth screenshotting.
- An AI software factory takes a planning document as input and ships production code as output, with no human reading the diff in between.
- The five levels of AI coding autonomy run from manually writing every line up to a 'dark factory' level where the engineer only writes the plan and the issues.
- A year-long dark factory experiment produced a live AI tutor app that was never read or edited line by line by a human.
- Coding agents, the language models underneath them, and the workflows wrapped around them are all improving at the same time, which is what makes full autonomy newly realistic.
- Mid-level autonomy, where AI writes most of the code and a human reviews everything, turns that human into the bottleneck for every merge.
- Full autonomy removes the steering wheel entirely: the human sets goals, and the system defines implementation, writes code, writes tests, fixes bugs, and ships.
- Proof-of-concept and idea-spiking work is now realistic to hand to a fully autonomous coding pipeline because language model costs have dropped so far.
- An open source AI software factory is being released that installs itself from a single prompt pasted into a coding agent.
- The factory runs on top of an existing open source workflow engine that turns plans into implemented, tested, and deployed code.
- The next stress test for the factory is video games rather than web apps, because games keep offering room for more complexity instead of hitting a feature ceiling.
Where AI coding autonomy realistically stands today
AUTONOMOUS CODING PIPELINES
A five-level model for how much control you hand a coding agent turns out to be a genuinely useful way to decide where your own workflow should sit today, not a thought experiment.
01A PRD Goes In, Shipped Code Out
- An AI software factory treats a single planning document as the entire input, chaining agents through planning, building, review, merging, and deployment before a human ever sees the code.
- The premise sounds risky specifically because it removes the review step most teams treat as the safety net for AI-generated code.
02Dan Shapiro's Five Levels of AI Coding
- The five-level driving-automation analogy gives a shared vocabulary for how much control you're handing an agent, from AI as a reference tool up to full autonomy.
- At the mid-level, where AI writes most of the code and a human reviews everything, the human becomes the bottleneck, since every merge still waits on a person reading a diff.
- Full autonomy means the human only sets goals and reviews outcomes, while the system handles implementation, tests, bug fixes, and shipping on its own.
03The Experiment: My Dark Factory and Dynachat
- A working, publicly used app can be built end-to-end by an autonomous pipeline without anyone reading a single line of its code, at least for a project of modest complexity.
- A successful low-stakes experiment doesn't prove reliability for complex, critical systems — it only proves the approach can work at all, which is a different claim.
05Why I'm Changing How I Teach
- Projects that ship as install-and-run tools tend to spread further than projects that only teach a process, even when the process-based content is well received.
- Admitting a strategy isn't landing the way you hoped, and adjusting instead of doubling down, is a useful habit separate from the specific pivot made here.
06Inside the AI Software Factory
- Building a general-purpose automation tool on top of an existing, actively maintained workflow engine avoids re-solving orchestration problems that are already solved elsewhere.
- An install flow that runs from a single pasted prompt lowers the bar to try a complex system, even when the system underneath is still early and rough.
07It's an Alpha, Come Build It With Me
- Releasing a system as an explicit early alpha, with a stated invitation to build it out in public, sets different expectations than releasing something framed as finished.
08The Projects I Need to Build With It
- Testing an automation pipeline requires real applications built alongside it, not just the tooling itself, because tooling without a genuine use case doesn't surface real failure modes.
09Why Video Games, What's Next
- Choosing a genuinely complex target application, one with room to keep adding features rather than a fixed scope, is a better stress test for an autonomous build pipeline than another simple web app.
- Pushing a coding agent to its limits surfaces the same reliability lessons whether or not you personally stay fully in the loop, so the lessons transfer even to more supervised workflows.
Terms worth knowing.
- AI software factory / dark factory
- A pipeline of AI agents that takes a planning document as input and produces shipped, deployed code as output, without a human reviewing the code in between.
- PRD
- Product Requirements Document — a written plan describing what should be built, used here as the single input that starts the automated pipeline.
- Level 5 autonomy
- The highest level on a five-level AI coding autonomy scale, where the human gives plain-English goals and the system handles implementation, testing, bug fixing, and deployment on its own.
- Workflow engine
- The underlying automation system that drives the multi-step process of turning a plan into implemented, reviewed, and merged code; the factory in this video runs on top of one called Archon.
Things they pointed at.
Lines you could clip.
00:00
“A PRD goes into the system and you get shipped code out.”
03:22
“There's no steering wheel anymore.”
04:11
“I never wrote or even looked at a single line of code for this application.”
11:59
“This might sound kind of silly, but what I want to build is video games.”
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 AI software factory, or what I've been calling the dark factory on my channel this year, is the ultimate evolution of AI coding. A PRD goes into the system and you get shipped code out, right? You create your higher level planning document, you give that to the system to split into individual tasks, it goes through building all of them, reviewing the pull requests, merging things, and getting everything deployed straight to production without a human looking at the code.
Now you might be thinking there is no way that could ever be reliable enough, especially if you're of a skeptic for coding agents in general and i get that but i've been hearing about companies left and right now that are legitimately implementing this kind of system for at least part of the development in their organization i've even helped a few businesses build out an ai factory system because here's the thing Coding agents as our harness, large language models, and our larger workflows are all getting better in parallel.
And so having this kind of system, it is becoming realistic. I know that might seem like a hot take for some of you, but I think it is legitimately productive now for any business to incorporate some kind of AI software factory, at least for prototyping things quickly and spiking product ideas. And let me tell you, I am betting big on this.
The next big thing that I'm gonna be pouring a ton of time into and focusing on with... my YouTube channel is building out an open source AI software factory, something that you can download and use immediately to start building anything. And I know this is a big goal.
I'm just getting started on this here, but I truly believe that the future of building software in general is getting to this level of autonomy. And so that's what I want to cover in this video. I want to talk about what a software factory really is, how I've gotten here, my own experimentation I've done this year, and I want to talk about what's next for this project.
And hey, if you've never really followed my channel before, you've just checked out a couple of videos, this might be the time to jump in. It is going to be a fun journey. I'm going to go on here with the software factory.
And don't worry, I'm not some delusional freak that thinks that AI is going to replace all engineering in a year. I just think that the software factory is legitimately going to become a big tool in the tool belt of most companies within the next few years. And we get to lead the charge together.
And just one more piece of food for thought here. Another reason I really want... to get invested into building a software factory is this forces me to make AI coding assistance as reliable as possible.
If I take myself out of the loop, I can't catch the small mistakes my agent is making. And so I'm just making myself a better agentic engineer when I push things to the limits like this. So if you follow along with me, you're going to be taking all the lessons I get from doing this as well.
So an AI software factory can also be called a dark factory. They're pretty much the same thing. In fact, I used to call it more of the dark Dark Factory.
Just because that's what Dan Shapiro calls it in the blog post he put out at the start of the year, this is what really got me interested in the idea. He covers the five levels of AI coding, basically the different levels of autonomy we give our coding agent. I've covered this on my channel a couple times before.
I'll probably keep referencing it a lot as well. You could turn this into a drinking game, like take a shot every time Cole shows Dan Shapiro's blog post. But I love this because it uses the analogy of driving a vehicle to help us understand the different levels of AI.
coding autonomy. So starting with you pretty much still write all the code yourself, you're using AI for a little bit of assistance down to level three, which is mostly what I teach where you want the most reliability possible. You have the coding agent writing all the code, but you're very involved in the planning and validation for every individual issue or ticket.
But then we get down to level five. So we have a console here, we still get to drive the vehicle with the higher level directions, but . All the smaller decisions are made completely by the agent at this point we send in our planning document and we get shipped code out and that is the dream for anybody as long as it's reliable enough now this big question is the dark factory reliable enough is something i started an experiment on earlier this year to figure out so i built a full dark factory that i documented building on my youtube channel through live streams like this one and a lot of youtube videos and i created a full application with it i call this dynachat my AI tutor.
It's an agentic chat application that has full access to my YouTube channel. And then if you're in the dynamist community, you can also search through my course and workshop content as well. So you ask it any question and then it's grounded in everything that I teach on the internet.
It's pretty cool. So it's not the most complicated application, but the important thing is I built this end to end with my dark factory. I never wrote or even looked at a single line of code for this application.
I'll have a link to this in the description. You can literally try this out right now. It's a live production application.
that I didn't build a single piece of myself. So I would say overall that this experiment was a success. But the thing is, this isn't a very critical application.
It's not extremely complicated. I would say I didn't really truly test the reliability of dark factories. And that is what I'm going to be doing next.
And this initial experiment that I did, I even gave you a skill to help you build your own version of it. So I'll link to a video right here where I covered that. Within my main skills GitHub repository, I have built...
the dark factory. So it'll walk you through creating from scratch the exact system that I used myself. And so that's a cool starting point.
But this is where I have to kind of just be transparent with you and share a little bit of a change in my philosophy for how I teach and bring tools to you. The sponsor of today's video is Data Impulse, the platform for pay as you go proxies for all your scraping needs for your agents and beyond. I'm building agents that involve scraping all the time.
In fact, I have one mid flight right now that's econ. I have a very simplified example of that scraper right here. So I'm pulling product information, but the problem is I'm using a single IP address.
I'm not using a proxy. And so what happens is you start to get rate limited or outright blocked. This is the first wall you'll hit with any scraping you do beyond just a handful of requests.
And Data Impulse is the solution to that problem. They have their own residential network with 90 million plus IPs for you to use for your proxy. And it's across 195 countries.
Plus there's no reselling. So all the IP addresses that you use here are going to be fast and reliable. So going back to my simplified demo here, I'm running the same scraper, but this time routing all my traffic through data impulse.
And so instead of it just going through the same IP address every single time, you can see that now we are using different IPs and I'm not getting rate limited at all. And all the traffic that I'm generating with my scraper, I have the dashboard here where I can see the individual requests and all the metrics over time. And I promise I won't get too in the weeds, but just to show you the code really quickly, all I have to do is use the same scraper with their proxy URL.
There are no other changes to the code and now all the requests come back from a different residential IP. So it just seems like a bunch of different people are interacting with the site from their homes. The whole thing runs on $1 for every one gigabyte, pay as you go and nothing expires.
And I can even hold a sticky session if I have to use the same IP address across a couple different requests. Data Impulse is the proxy solution. for any kind of scraping solution that you're building, like rag agents, price tracking apps, whatever it is.
I'll have a link to them in the description. So I hope you don't mind. I'm just gonna get vulnerable for a little bit here.
So for the lifetime of my YouTube channel, I've always had the philosophy, I wanna teach you how to build something yourself instead of just giving it to you, right? Like . And I know a lot of you appreciate that with my content, so I'm not going to ditch that.
But now I'm rethinking things. I want at least... of my content to be i'm actually going to just give you the tool to run right out of the box like before it's always like here i'll give you the skill to help you build your dark factory or i'll teach you how to build your own second brain from scratch but i gotta be honest i've always wanted to have an open source project like open claw with almost 400 000 stars or hermes asian with 240 000 stars deep sea carnus that's already at 209 000 stars like i've never had something go viral like this i mean the closest i've gotten is with archon which fantastic tool i'm still pouring a lot into this.
And it's great, but it's not a 200 ,000 star, 300 ,000 star project. And I think the reason why none of my stuff has gone sort of viral like these other tools is because it's not something incredible out of the box. It requires you to build out your own process or to really create something from scratch because it's always been my philosophy, like I said.
But now I think it's my turn to build something that you can just use right away super easily and get a ton of value with it. So trust me. is not going anywhere.
My philosophy still stands most of the time, but I also get it. Like sometimes you just want something that you can just download and start using right away with barely any setup. And so that is what I'm doing now with the new AI software factory.
I'm taking everything I've learned from helping companies implement systems like this, from my public experiment I've been covering on my channel, all the lessons that I've learned the hard way. And I'm building that now into a solution that you can just download and use right away. So obviously it's quite a pin.
opinionated how I like to build software, what I've ingrained into this. And I actually even have this running Archon workflows under the hood. So when you install it, it installs Archon to drive all the processes that take your plans, turn it into written code that's deployed.
And of course, there'll be dials you can tweak to really customize things and make it your own if you want, if you subscribe to my usual philosophy. But the point is, there's now the option where you can take this prompt at the top of the readme, send it into your coding agent, it'll ask you a few questions and then it'll install everything and be immediately ready to accept any PRD and just build things end to end for you.
Just working right out of the box. That's not my usual thing, but that's what I've got here. And like I said earlier, I just started building this.
And so the point of this video is not to give you a tutorial to set this up and run it. It's not necessarily ready for that. I mean, if you want to start the alpha of the software factory, you can go ahead and try it right now.
Send this prompt into your coding agent, but More why I wanted to make this video is to show you what I'm working on and invite you on a journey to build a software factory. So come along with me as I build this out as a better and better resource for you.
And we'll see together how reliable we can make this kind of system. I'm decently skeptical that we can use a factory to build anything. In fact, you might even be a little bit more skeptical or less.
I don't really know. The point is we're going to try to find that bar. And I think we really can.
Like if you need a certain level of reliability, you still want to be in the loop. lot more, but for certain kinds of tasks or things you're working on, you can just give it over to a software factory as long as we put enough effort building out the harness, which is exactly what I'm going to be doing. And like I said at the start of the video, there's certain things that I already know you can use this for, like proof of concepts and spiking product ideas.
Like literally using your coding agent to just build out entire things to test ideas is realistic now, especially because of how cheap certain large language models are. So you can be very cost effective in building out a ton of code just to try things out. And trust me, I've been spending an insane amount of time testing this software factory as I'm building it out.
So I create a ton of different software factories from scratch, like the one you're looking at right here. And I do that over and over again as I iterate and just continue to refine the main open source project that I'm sharing with you. Because it's been a lot of work to take all the lessons from the initial experiment I did and the companies I've worked with and...
turn that into the single open source project. And so I've been spending a lot of tokens building this out and we're just getting started. Now, at this point, I don't have one of these test repositories to share with you yet, but that's one of the next big things I'm going to be working on.
Obviously, as I build out the software factory as the main repo for you, I also need real applications I build alongside it to prove the factory is working well. So I'm going to be putting a lot of effort into that. I've been thinking a lot about the kinds of applications I want to build too.
So the AI Tutor application that you're looking at right here, what I built for the last experiment, this is great. It's practical, it's deployed right now, it's a great resource for you to use. But when I really push the limit of software factories coming up here, I need something a lot more complicated.
And so honestly, this might sound kind of silly, but what I want to build is video games. Video games are fantastic for testing complicated build -outs because applications like this, there's only so much you can do before it just... becomes a feature bloated mess but with video games it feels like there's always an opportunity to build in more features do something more elaborate and more fun i think it's going to be fantastic so i'm not going to do anything web -based like this i know that these might be a bit more practical for you know resources for my channel but video games are fun for me like i grew up playing video games i still play them and i think it's just going to be a great way to push the limits of software factories and so all that to say i've got a lot of cool things coming up on my YouTube channel where I'll be building out the software factory, making it better and better for you, showing you the different projects that I've been building with it, and of course, teaching you all the lessons and best practices I learned as I'm building out the harness underneath, talking about how I use Archon and those workflows to drive everything, best practices to make things more reliable.
I do also already have content on my channel for how to build a dark factory. So I'll link to a video right here where I cover all of the concepts in this diagram. So A lot of golden nuggets that I've covered already, but I just want to keep hitting on this as I work harder and harder on the project here.
And honestly, the most important thing to mention here is that everything that goes into building this software factory is also going to teach me and you how to just use coding agents better in general. Because if we are taking the autonomy to this level, we better make things reliable, but that's also going to help even if you're more in the loop.
And so there is just a lot to get out of this journey and following along with me. if you appreciated this video and you're looking forward to following along as I build a software factory and all the AI coding lessons from that, I would really appreciate a like and a subscribe. And with that, I will see you in the next video.
The Hook
The bait, then the rug-pull.
A planning document goes in, and shipped code comes out the other end, with no human reading the diff. That's the pitch for the AI software factory this video opens with, and the rest of it is the case for why that's no longer a hypothetical.
Named ideas worth stealing.
02:37model
The Five Levels of AI Coding Autonomy
- Level 0 — Spicy Autocomplete: AI serves as a reference tool or enhanced search; the developer manually writes all the code.
- Level 3 — Developer: the AI generates the majority of the codebase; the developer reviews everything the AI does, acting as the bottleneck for verification.
- Level 5 — Dark Software Factory: the engineer manages the goals and the system, not the code; the AI defines implementation, writes code, writes tests, fixes bugs, and ships.
A driving-automation analogy (borrowed from Dan Shapiro's blog) for how much control a developer hands to a coding agent, running from manual coding up to a fully autonomous factory.
Steal forframing any conversation about AI-coding adoption on a shared vocabulary for how much control to hand an agent
How they asked for the click.
VERBAL ASK
13:56subscribe
“If you appreciated this video and you're looking forward to following along as I build a software factory and all the AI coding lessons from that, I would really appreciate a like and a subscribe.”
Direct, low-pressure ask placed after the full arc of the video, tied to following the project's ongoing progress rather than a hard product pitch.
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.
