This Agentic Coding Setup is Ahead of Its Time (42k Commits in 45 days is ABSURD)
January 8, 2026
Parker breaks down an agent-driven coding setup—the agent scripts folder—and shows how centralizing configs, prompts, and tooling can dramatically accelerate shipping and consistency across projects.
What is the Agent Scripts folder?#
- A central hub of configurations, docs, and commands that point to all your other projects.
- Prevents rewriting the same setup in every repo; you point new work to the agent scripts directory first.
- Acts as a single source of truth for workflows, tooling, and “dashboards” you’ll reuse across projects.
Why this matters for engineers#
- You become an “agent operator” rather than forever tweaking configs in isolation.
- Version-control for your entire tooling stack: you can evolve the agent scripts once and apply it everywhere.
- Easy access to an ops-oriented playbook: shipping, debugging, and deployment workflows are documented and referenced in one place.
Core structure you’ll encounter#
- Central barrel/entry point that new folders reference first (reads agent scripts/README, etc.).
- Docs and slash commands baked in for quick access to common tasks.
- Pointers to the dotfiles and root-level scripts in each project.
- A committer script and a suite of tools (CLI helpers, prompts, etc.) that automate routine tasks.
Code example (high level)
/agent-scripts
/docs
/commands
/tools
/skills
README.md- Your actual projects then point to this directory in their first-possible place to read agent.config or agent.mmd, so you never duplicate setup.
Key components Parker highlights#
- Ghosty: the environment context Parker uses (light/dark mode, navigation around home/projects).
- Barrel file layout: a single entry point that routes to per-project configs.
- Committer script: smartly handles commits when you’re shipping lots of small changes.
- CLI tools (e.g., Nano Banana): handy utilities built into the workflow.
- Skills framework: a modular way to package capabilities (see below for structure).
- Refs and scripts: references to other tools, APIs, or open-source projects you rely on.
The “skills” concept and how to use it#
- Skills are folder-based modules; the folder name is the skill name, and the file inside uses ALL CAPS with front matter.
- Front matter helps package metadata for quick reading by the agent.
- You can chain or reference skills (refs) and even script them to be called from the main workflow.
- Example pattern:
- Path: skills/create_cli/CREATE_CLI.md
- Content starts with front matter, then a concise description and usage notes.
Code block (skill file skeleton)
/skills
/create_cli
CREATE_CLI.mdContents (example):
---
name: Create CLI
description: Scaffold a ready-to-run CLI with argument parsing
frontmatter:
category: CLI
author: Parker Rex
---
Usage:
codeex$ CREATE_CLI
...- This packaging lets you reuse CLI scaffolds, UI/UX prompts, and other capabilities across projects without reimplementing.
How Parker uses it in practice#
- He integrates this with codecs (the local workspace) and uses it as the source of truth for project scaffolding and ops tasks.
- You can point new repos to the agent scripts folder so every project inherits the same foundational setups.
- Over time, you replace or customize pieces specific to your stack while keeping a consistent baseline.
Real-world workflow benefits#
- Ship more, faster: fewer repetitive setup steps across projects.
- Simplified stack maintenance: you swap in better tools (e.g., Opus or Codex) when needed without ripping out your entire workflow.
- Clearer handoffs: new team members can learn the system by reading the agent scripts and the linked docs.
Practical prompts and templates you can borrow#
- Feature analysis improvement: pulls patterns from open-source work to guide UI/UX, hooks, and logic.
- Refactoring opportunities: generate an extended list of improvements (often 30+ items) and prioritize them.
- Spec interviews: prompts to surface design questions, poke holes, and clarify requirements.
- Spec-to-prompt integration: use tools like “Ask user questions” to flesh out missing spec details.
- Lists and knowledge sources: build Twitter/AI power-user lists to surface best practices and prompts.
- Park Rex has extracted prompts on parkrex.com (Writing tab, press E) you can reuse and adapt.
What this looks like in practice:
- Build a Shorts Factory prompt set to design AI-assisted educational videos.
- Use a "video gen" prompt to storyboard, fetch a CapCut-like open-source reference, and orchestrate browser APIs for audio/video handling.
Quick-start actionable steps#
- Create a dedicated agent-scripts folder in your workspace and point new projects to it as the primary config source.
- Start by copying a minimal subset of Peter’s repo and tailor it to your stack (remove things you don’t need, keep the essentials you use daily).
- Define a few core skills (CLI starter, simple automation, a UI prompt) to prove the workflow and then expand.
- Document your workflow in the agent scripts README so teammates can onboard quickly.
- Experiment with the built-in prompts for feature analysis, refactoring opportunities, and spec interviews to bootstrap automation in your projects.
Takeaways#
- Centralized agent scripts unlocks a repeatable, auditable, and scalable workflow across multiple projects.
- Treat your tooling like code: version, review, and evolve it rather than duplicating configs.
- Start with small, reusable SKILLS modules and gradually expand to cover your common tasks and workflows.
Links#
- steipete/agent-scripts - Peter Steinberger's open source agent scripts repository
- RepoBar - macOS menu bar app for GitHub repository monitoring
- Ghostty - Fast, feature-rich terminal emulator with GPU acceleration
- OpenAI Codex vs. Claude (agents/AI tools)
- Claude Opus (execution/automation tool)
- parkerrex.com prompts page (Writing tab; press E to view prompts)
- Anthropic and related AI prompt tooling (as mentioned for follow-up prompts)
Transcript
Hey there, I'm Parker Rex. I a le tech for a startup that sold for 23 million bucks. After that, I went all in on engineering, started a consultancy, done a bunch of different stuff. And so, I want to talk about a trend that I find very, very interesting. If you've been watching this channel for a while, we're usually ahead of the game on a bunch of the trends. We try to cut through the BS that is most of the internet on AI stuff. So, it's called an agents scripts folder. And the reason I'm on this guy Peter's page is because he has an open source one that's really good. 42,000 contributions in the last year's bananas. Now, you're probably thinking, "Oh, he just has this automatic thing that just goes and changes a number or something." No, it's not the case. These are all these current projects and I have a few of them installed myself. Repo bar is really nice. It shows you all of them in one place. helps you solve a new problem that's happened which is you're shipping a lot of stuff and you have a lot of open source work. You want to be able to have one spot for it. That's what that does and you can just go and explore the on your own. But the agent scripts is essentially a set of configurations that you have as a central place that are served as pointers to every other project. What does that mean? It means you shouldn't be rewriting these configs everywhere that you go, right? And so by having this in one spot, it's super helpful. It includes documentation around the workflows that you're doing. Includes slash commands as well. So this acts as a pointer to the dot files that are at the root of your project or your home screen. So if I open up Ghosty and sorry about the flashbang, we're going light mode. If I go into my home and I do a cd intocodex and let's make this big shift. Why is it working? There we go. If I do an lsla, you'll be able to see what's in here. And so this is really where codecs gets served from. But you'll notice that at this other location, the agents scripts folder, you have pointers to the actual ones. And why that's important is it basically allows you to have version control of everything that you're going to have across every single project for leading your agents. Now, why does that matter? Because you can give it a bunch of different tools. So, if I go into my projects and then I go into agent scripts, you'll see I have taken his repo down and I've started to modify it to my system needs. So every single time that you'd then go use a new folder, a new repo, new project, in the first line of it, you would point it to this directory saying before anything else, go read the agents.mmd from agent scripts. And it's really nice because you just won't repeat yourself. So if I go into let's say the actually let's just take a look in here you can see docs and he has a bunch of them in here that are specific to different technologies as well as how to do certain things. So releasing a Mac product there's a lot of steps that are involved with that. And what's nice here is this is almost like using the syntax of skills but it's really just something that like you need to refer to from an ops perspective like shipping or debugging a Mac OS release. You're going to use this thing called sparkle, which is in ghosty, funny enough. And then you had have some other things in here. I've added some that are web related. You have your slash commands in here. And that's just a barrel file. But you'll see that this little read when thing is really helpful. Now, I will be honest. I looked at this at first and I was a little overwhelmed, but I went and I just changed everything out that I found was specific to his projects and then tore a bunch of stuff out that would make no sense for me to have and started having it be a pointer to my other projects or started pointing other projects to this. And little things like this committer script really nice. When it goes to commit work, it just figures it out. So he has different scripts that are in here. He has a CLI for Nano Banana and how to be able to use that which is pretty interesting. You have the skills. So this one's very interesting where skills are similar syntactically both in codeex and claude, but there's just different ways to call them. So you use a dollar sign in codeex, but reality is you should have them in one spot to be able to be referred to. And if you didn't know with skills, the syntax is basically name of skill is the name of the folder and then skill in all caps with some front matter at the beginning so it's easy to read and then a bunch of information nice and packaged up. But you can also have refs and you can also have scripts. So that's what this one is. You can have a skill for creating a CLI. This guy starts with CLIs a lot of the time and then we'll add a front end which is really nice. I used that yesterday to make a CLI and it just worked. So you can see a bunch of stuff in here. It's crazy. This is absolutely the sauce right here is figuring out how to operationalize these agents so that when you go to work on anything ever, it has all the tools. And this is going to be one of those things similar to dot files with invim where you're going to be constantly tweaking it, figuring out better ways to use it because engineers are now agent operators. This is really what it this meta skill that's exploded in popularity because people know that this is the alpha. So that's what I would do is I would go and grab this. I think I covered a bunch of the highle stuff. He has a tool reference in here. These are all tools that he's made. So check him out for sure. He has a really good blog post as well. It's something at inference yeah shipping at inference speed. And so this was what changed my opinion on it and made me focus on essentially first of all using codecs a lot more but then making the workflow better. So how did that come out in practice? I shipped a lot more stuff. I got a lot of the projects that I've been working on for a long time across the finish line. I got to change out and simplify my stack on some other side projects where I wanted it to eventually be on convex because it replaced a bunch of stuff, but I didn't feel like going and doing that annoying work and codeex just crushed it. And so a lot of the stuff in here is worth reading as well. I pulled some of the prompts out. So if you go to my website parkrex.com, you just hit the E button on desktop, you'll be able to see under the writing tab that I've extracted a bunch of these different prompts. and I'll go through them real quick on how to use them. So this first one, this feature analysis improvement one, there's so much open source work out there that has figured out the patterns for you. So take advantage of that. Now you shouldn't like copy paste their whole thing, but you can pull out patterns from how they do the UIUX, how they do the hooks and the logic, sharable code, whatever, and instruct your agent to do that. So in my case with this shorts factory that I have, what it does, the problem it solves is I want to make educational thriller science pop or popular science videos that are totally animated using AI from start to finish, researched, all that. How do you do that? So it takes a lot of steps, but how I used that prompt was towards the end, I needed some video editing features. Oops. I needed a like essentially something like Cap Cut where I could have a player and a play head and all this stuff. And I'll just see it wants to start. I got a Docker going, but if we start it, you can get a better idea. What I did was for this particular prompt, it is started. Cool. It says our feature doesn't work well. And so our feature was the play head and a bunch of stuff around that. Look at one that's built in. And then I have this AI refs open cut which is cap cut open source analyze how it works yada yada. You put that in there. Use voice dictation to put in a bunch of other information. It's helpful for it to know and then it'll go and learn basically what what is it that's the difference between your implementation and their implementation. And so I have this refs in here with both CAP which was for figuring out how to connect devices properly working with browser APIs to understand basically if you have audio recordings on different screens you have to like open and close those connections. I didn't know how to do that and so I put that in there and I put this one in there and then for the video gen I grabbed a Google example and you just say go and analyze it and you speak to codeex in a way where it's conversational so that you yourself can also understand it because you're going to be orchestrating these things at the end of the day and you're far more effective when you know how to do it yourself because it's just better that way and your brain won't atrophy as much. So that's how that one works. it'll end up coming up with a really good plan and then you can iterate on that and go and execute. The next one is refactoring opportunities. I took this one from that same guy Peter. It's basically like an ode to all the amazing kind of aentic stuff that he's done in this video. But now you've read so much code. What are the biggest opportunities to refactor and improve the code base? Give me an extended list of what makes sense. This one's fantastic. And this could also be really good if you were just in an open source project and you were looking to contribute. You could go in and ask a bunch of questions about how things work and then do that. It'll give you about 30 things which is crazy. And then actually executing on those you can what I found helpful is I'll have it write them out to an AAI slash plans and then I can say maybe here's a bunch of them. I could be like here's all these opportunities for UIUX improvements and there's all the different phases and if there's how many phases are in here? four phases with 30 to-dos. I'll ask if you can put those in a prioritized manner where they can run on multiple tracks such that there's no cross cutting. And that then allows you to open up three different cloud codes or whatever you want to use. But I use Opus most of the time until I or sorry codeex most time until I run out. And then I use cloud as well for execution. Codex is slower, better at async stuff, and Opus is obviously very great, too. So, that's how I use that one. Then, when it comes to these spec interviews, if you're following the folks on Twitter, there's a bunch of great people that work at these companies, and you can just add them to lists. So, I recommend making lists and then like for instance, I have all these lists and yeah, it's I had stupid politics ones when I was trying to figure out what the election was doing. But anyways, in here I stack in the people that are interesting where it's oh AI power users. These are people that have I think it was the top 20 token usage. They had a list of them. So I just put them all on a list and you can find interesting stuff. The cloud code folks, you can go and follow Boris and then see, oh, who does he follow? And then just type in Enthropic and you'll see everyone and then you can go and make a list for Enthropic and then you find out good prompts like this. So, what this does is it acts more like a product manager than anything and a tech lead or I guess like an architect. So, it'll ask a bunch of questions about the UIUX and poke holes in it. It's asked me like 50 questions before. Basically, the weaker that your spec is, the more stuff that it'll find. And I find that the ask user question tool is really good in Clawed out of the box because it's just a tool that's set up. So what the codeex team did was they actually made this one which is ask questions if underspecified and then you have to go and make the skill. So if you want that skill you can go and what's his name? Hang on. I just saw this is the guy from anthropic that did that ask user questions tool and then the codeex folks pretty much copied the exact same thing. If I go into where is it? opening eye and then where's our guy Tibo. So T-Bo if you go to his and then you go to media and you scroll down you will be able to quickly find something in here about ask user questions and then you post the skill right below it. So these guys are definitely worth following. I think it is here. Yeah, there it is. Skill file. Bam. So you can just take that and then if you're not familiar with how to do that then I can just show you really quickly before I wrap this video. So let's go back all the way. Yeah, there. And then codeex. Yep. In skills and right I just saw it. There it is. So you'd paste it in there and then when you want to call it, you just type in codeex dollar sign and then all them right in there. So very cool. Very cool. I just wanted to catch you guys up on some of the stuff that I found interesting that I think is worth talking about. I highly recommend going and making a agent scripts folder. I think you'll learn a lot about just things that you're doing repetitively, whether it's in the same project or just managing your computer. But if you like this video, do the thing that you're supposed to do, which is like it so that more so that more people can find it and subscribe to the channel. and I will see you in the next