Loop Engineering in 9 Minutes

I saw a couple posts that went viral of the creator of Claude code and he mentioned, I don't prompt Claude anymore. What I mostly now use are loops. I create loops. They do the rest of my job. And it's a really interesting concept. I think for a lot of people that are potentially new to this type of thing, it might seem a little bit out there, especially if you have the type of role where it might be a little bit more hands-on. But the interesting thing with loops and what I'm going to be showing you in this video is you can use them in a number of different ways. There's really a spectrum in terms of how. you can use them. I'm going to touch on some ways in terms of how I personally use them. Now, an interesting post that just came out with the release of Fable 5 that came out yesterday was from Lance Martin. He wrote a really great X post, which I'll link within the description of the video. And basically the idea around loops is there's different levels of complexity with how you can actually approach this. There are some projects like auto research from Andre Carpathie, for instance, where what it does is it will train a model by performing different experiments at a set interval. It will look and see what is actually working, what isn't. And for the things that work, it will progress it along its way. And this type of idea can really be applied in a ton of different ways. He mentioned within the post parameter golf is an interesting way to actually go about this. Another way that you can actually think about this is even broader on tasks that aren't just even programming related. One you can use both across codecs as well as cloud code is the concept of a goal. Now a goal is essentially a long running task. is a very similar to something like Ralph Wiggum's where instead of actually having to prompt the model again and again to go and continually work on a problem this will automatically have the mechanisms built into the harness where it will carry on its merry way until it actually gets to the goal what you can do with this and this is probably one of the easiest developer experience or just user experiences in terms of how you can set this up is within cloud code or codex you can simply run goal and then you can describe what you want to do now what you can do is within Claude code specifically is they do also have a loop feature. For instance, what I can do within here is if I slash loop, you're going to be able to put in your interval in natural language. You could do it just like this loop every five minutes, and then you can describe the outcome of what you want to do. Now, the one thing that I particularly like about loop is oftentimes if you're doing really exploratory work, or if you're just experimenting with different things, maybe it's a greenfield project and you really want to leverage. all of the capabilities that's built into the model and what it's actually able to do is oftentimes what I like to do is I'll run a loop with a particular broad open-ended question. And the way that you can think about this is almost think about it, handing it off to a really ambitious junior engineer. It might go off in a direction that doesn't necessarily make sense, but oftentimes it might not necessarily even make sense to you that it isn't a good approach to try until you actually see the results back. And as things get progressively cheaper with being able to try out different new ideas and code just generally is becoming cheap you can do much more ambitious things the thing that i like with loop is i can say something like okay every five minutes let's go and let's explore these types of things that i'm trying to do within this project and what it will do is it will set up a cron job where within this session it will run for the duration of however long you specify so i could say loop for five minutes for the next three hours to help me explore this part of my project and you will know there's that deterministic aspect where it will trigger the cron job, where it will actually make sure to have that event fire within the context of the thread of the LLM. It's very similar to goal. I probably use this one a little bit less than goal because I find goal is actually very reliable. I've had goals that have run for multiple days for particular things that I've wanted to set up. So for instance, I've wanted to set up parsers for very complicated sets of documents across a number of different websites. And when I have just been experimenting with this goal feature, to see how far I can actually push it. You can actually have these run for literally days. If you have a task that is bounded. And one of the things that is really helpful for these types of tasks is if it actually has something to verify its worth. For instance, if you're trying to map something that's particularly big, that actually has a verifiable aspects about it, where you can run unit tests or whatever on it, those can be particularly good tasks for how you actually verify this. increasingly how i've actually been preferring to set up more long-running automations that are very very helpful to my day-to-day work is within environments like codex cloud code as well as cursor there's a ton of other ide's as well that have a similar feature and they all more or less take the same shape essentially there's going to be an automations tab or a similar name for it within all of them in the top left hand corner typically and then you're going to have a layout something like this you're going to have the current automations that you have and then you're going to have the paused automation Now, these sort of came and went as a little bit of an announcement. And then I actually didn't see a lot of people focus on these, myself included. I didn't actually leverage these a whole lot initially when they came out. I didn't actually think about how I could consistently leverage these things in my day to day. As I started to have one that does something useful every single day, I've started to think of others that actually just push along a lot of the minutia that I would have otherwise had to deal with. Now, with my YouTube channel, I get a ton of emails and I don't really have a whole lot of time. manage different requests or just people reaching out with different questions for various things so what i have is i have a automation where essentially what it does is i say look through my inbox and go and put everything within my inbox on a linear board that i have and what it will do is every day it will check the inbox it will check the linear board and it will essentially see what are the highest priority things that i need to action now how this can also be used and this is a relatively simple one but you can also have this be updated in a way where you can leverage the internal files that you have within codex or cloud code where you can say at this interval create an agent.md for this project or create relevant skills on all of the things that i might have done over the course of the week now in terms of some other ways that you can actually leverage this is if you have a project you can have it scanned for security vulnerabilities and those types of things at a particular cadence you know another way that you can leverage this is as a form of memory now there's been a few posts that i've seen around where people have discussed an interesting idea where you have your LLM or your agent actually go through and essentially dream about all of the different things that had occurred over the past day. And in a similar way for how we store memory is what we can do is we can have these systems create a synthetic representation in smaller, more coherent and accessible and progressively disclose ways to the model. Essentially a way that models can actually programmatically load up into their context window and know where to look when about particular things. And this is the idea around continual learning, which I think we're only really going to be hearing more and more about over the coming months. It's really an area of focus for a lot of people, a lot of researchers, because what you can do with continual learning is essentially once you have a system that can learn and doesn't just need to be stuck with the pre-training of the model and the harness, and it can actually take and explore different tasks within the world and ultimately build on those different things over time. That is a system that can progressively get smarter over time and be more useful across a ton of different domains so there's a ton of different ways in terms of how you can leverage loops it can go from tasks just helping you push along within a slash loop it can be a goal that can run over days and days across different harnesses or it can also just be for tasks and things that you do repetitively every day every six hours once a week all of those different things if you think about it of all of the different tasks that you might do if there's something that you are repeatedly doing and it is computer, there is a pretty good chance that you're going to be able to offload a lot of these tasks to these types of automations. And the one thing that I do want to mention with these types of automations is you don't need to have it go end to end. You can still be in the loop. So for instance, when I have it draft my emails, I don't personally have it send emails. And oftentimes it doesn't actually get all of the context right in terms of what I should do. I just use it as a helpful assistant, but over time you can actually have it improve those prompts in and of itself for how you're actually leveraging these things. It's definitely a really interesting topic of discussion about how you can leverage these loops, but otherwise I'm curious your thoughts. Are you spending your time? like Boris Journey? Are you moving away from prompting? Are you starting to think more about building loops rather than actually prompting systems? Leave them within the comments below. Are you leveraging loops more? Are you actually spending less time prompting the system and now actually scaffolding out more automations and thinking things at a much higher level in terms of all these different loops that your systems and agents manage? Let me know how you're leveraging things like loops or automations within the comments below. But otherwise, if you found this video useful, please like, comment, share, and subscribe. subscribe otherwise until the next one

Stop Prompting, Start Building Loops: Goals, Automations, and Long-Running AI Workflows

The script discusses shifting from repeatedly prompting LLMs to using long-running “loops” and automations, inspired by viral posts from the creator of Claude Code and an X post by Lance Martin following the release of Fable 5. It outlines a spectrum of loop complexity, from experiment-driven systems like Andrej Karpathy’s Auto Research and ideas like parameter golf, to practical features in Claude Code and Codex such as “goal” (a task that runs autonomously until completion) and “/loop” (interval-based runs like every five minutes). The speaker explains using goals for multi-day tasks like building parsers with verification via tests, and using IDE automations to handle repetitive work such as triaging a YouTube inbox into a Linear board, generating project docs, scanning for vulnerabilities, and creating memory-like daily summaries tied to continual learning.

00:00 Why Loops Beat Prompts
00:43 Loops Spectrum and Examples
01:18 Goals Long Running Agents
01:56 Claude Code Slash Loop
03:34 When to Use Goal vs Loop
04:18 IDE Automations Tab
05:11 Daily Inbox to Linear
06:00 Security Scans and Memory
06:44 Continual Learning Future
07:49 Human in the Loop Tips
08:12 Wrap Up and Questions
08:45 Like Subscribe Outro