Parker Rex
All videos
Parker Rex Daily

I completed 70 Coding Tasks for $1 Using Augment and VsCode (absurd)

June 3, 2025

Watch on YouTube@parkerrex2

I pushed a dense, hands-on flow using Augment and VSCode to tackle around 70 coding tasks for $1. Here’s the core setup, what actually worked, and how you can apply this pattern.

Workflow snapshot#

  • Augment + enhanced prompt button = fast, context-aware prompts that pull in paths, dependencies, and file trees.
  • Cursor Taskmaster integration = turn AI-assisted work into a deterministic sprint system.
  • The aim: a simple, repeatable flow that scales with teams and agents, not one-off hacks.

From idea to sprints#

  • Start with your real-world project (VI2 private network) and map it to a clean task structure.
  • Use a PRD-style plan inside the task system to keep what you’re building tangible and shippable.
  • Break work into sprints with a max of 10 tasks each to keep momentum and avoid token bloat.
  • Keep a lightweight backlog/doing/done pipeline, plus a deferred tag for mid-course pivots.

Architecture and tool decisions#

  • Focus on long-term tool viability and LM compatibility. Pick stacks and patterns the AI can roam across easily.
  • Favor patterns that enable remote agents to own codebase parts and run autonomous sprints.
  • Build around a "domain-based" structure with clear endpoints and tests so the AI can respect boundaries and stay productive.

Implementation notes for VI2#

  • Domain map examples: admin, announcements, events, GitHub O, feed, health, learning, members, projects, prompts ID, Stripe, Sentry, users, learning goals, notifications, onboarding, profile, etc.
  • 35+ API routes were scoped and tested; most functionality ended up covered by tests.
  • Architecture doc and diagrams (Mermaid) helped keep the AI aligned with the intended structure.
  • Tests-first approach: prompt the AI to write tests first, then implement code to pass them.
  • Integration points: GitHub models API for prompts (read-only in authorized repos), potential TRPC for future mobile app, and CI/CD considerations.

Testing and prompts workflow#

  • Write tests first, then implement code to pass them.
  • Prompt strategy examples: describe desired test coverage, then let Augment generate the test suite.
  • Ensure terminal/stack compatibility (ZH partially supported; uses Docker-friendly setups).
  • Use environment variables and code scaffolding prompts to keep the AI anchored to real-world constraints.

Results at a glance#

  • 13 concurrent threads, 7 sprints completed.
  • Roughly 70 tasks tackled via Augment/Cursor-driven prompts.
  • Cost: about $1 in augment credits for the entire run.
  • Clear progress signals (milestones, deferrals, and sprint boundaries) kept the work focused.

Q&A and community notes#

  • Where to learn more about folder structure and architecture decisions: dig into the architecture doc and the sprint/task structure examples.
  • Augment’s context engine is highly valuable for large codebases; the quote: “augmentation’s context is uncontested” reflects that approach.
  • Practical tips shared in comments: consider Claude Code or other cloud-backed prompts if you’re price-sensitive; the core pattern is the same—clear tasks, good prompts, and strict sprint boundaries.

Next steps and future work#

  • On the main channel: a deeper dive into tool selection, patterns, and how to pick a stack that scales with AI-assisted development.
  • Potential additions: remote agents owning codebase parts, wiring up TRPC for mobile later, and refining CI/CD around AI-driven work.
  • The ongoing goal: maintain a lean, repeatable workflow that balances AI power with human oversight.

Actionable takeaways#

  • Leverage Augment’s enhanced prompt feature to auto-collect context (paths, libs, file trees) for better continuation prompts.
  • Structure work as sprints with no more than 10 tasks; keep the PRD + taskdocs visible to the AI.
  • Write tests first, then prompt the AI to implement code to pass them.
  • Keep a clear Taskmaster object (name, sprint, status, environment data) to track progress across agents.
  • Choose tools and stacks the LM can learn and navigate; plan for remote agents to own code segments over time.
  • Use lightweight, endpoint-focused architecture docs and Mermaid diagrams to anchor the AI’s understanding.

If you want a tighter breakdown of the exact folder structure and the sprint templates used, I’ll run through a focused teardown in a follow-up.

Transcript

What if cursor taskmaster and augment had a kid? That'd be biologically challenging. But anyways, let's talk about it. This is the workflow that I am landing on. I am slowly diverging to which is the best of all the tools that I've been using by picking and selecting and keeping it simple. So, I want to talk about where I started, where I ended, and I will just show you. Let's see. We encountered an issue. I've been just loving this enhance prompt button. If you haven't used augment, the enhanced prompt button is the new cursor tab because it will be not only really helpful with continuation prompts, which is you finish a thread, it warns you that the quality is going to diminish and you try to bring that context to the next, which is automatically built into cursor. But it just makes it better. It I could say, "Hey, do this." And it will go and grab all the relative paths. grab all the relative libraries, dependencies, file trees, whatever, and it'll figure it out and make a great prompt for me to continue on. Now, I'm deep in this thread. You can see it's having a little bit of trouble, but it just finished this sprint. So, at least that's sprint task one. But what I've been thinking a lot about with all these advancements that keep coming out are how do you pick the tools that will last the test of time and how are they closer to the big provider companies. So there's typical setups like everyone's always trying to find when they're building stuff. And I'll go into this in depth on a new video on my main channel. It'll go out in a couple days. But you need your tools, right? And so you have your IDE. You need something that's going to solve for context. You need something that's going to solve for your prompts so you can actually get yourself through the AI software development life cycle, which is just a software development life cycle using AI by the way. So you need to know all those steps. And then lastly, you need to have some way of managing your work streams or tasks. But it's not just tasks. So a workstream is a PRD. It's a flow that you're following in order to accomplish the new feature or the new project. And I was talking about this earlier on our Discord with JRON, one of our members, and I figured I'd share. So what I did was I always just try to map things to what would you do if you did not have AI. And so in my case, I have this VI2, the new platform that I'm building for our private network. And it's got a lot of stuff in it that comes from school, which is where I launched the beta, the first version of it, but there's a lot of things that I need to add. And I'm trying to find the best patterns for the codebase such that it's really easy to do. And I hit a wall when I first built this where I chose a stack that the LM did not like. Now, if you're in the community and you're watching this, sorry, you have to hear me again, but I chose Tanac start. Long story short is it was like a 3 week spike for me to learn it and then for the AI to not learn it. And unless you're going to go and build your own pipeline for ingestion and dealing with like state machines and all this stuff, which would not make sense, just use augment or use cursor or whatever cloud code. Since you're not going to do that, you need to pick a stack that works. And then also the architectural patterns. So, I wanted to pick a stack that also had an architecture that LLM could jump around in a lot because I want to set up remote agents in a couple days or even maybe tomorrow where they'll own certain parts of the codebase and I'll have sprints that are assigned to different agents. So what I've done was I looked at the workflow that I would do in real life and I thought okay let's just use agile as an example and then we'll pair an agile workflow with the best parts of taskmaster which is having a very deterministic way of tracking your tasks. So I went in and I went to Taskmaster and I just jumped into their GitHub and I just asked a question to our dear friend Mr. Copilot. I said, "Hey, how how are you doing?" No, I said, "How are tasks laid created? What's the data object look like?" And look, he's following me around. That's crazy. So, it'll go and it'll find the relevant information in here. And then now I know what the structure of that looks like and I can compare it and compare and contrast it to what I have. And then I'll use that as part of this workflow I'm going to talk about in a second. So when I look in here, it's going to show me boom, this is what the object looks like. So there's some stuff in here that I like. So I yoink that. And then I realized, okay, now I have task structure, but like I said, there's a lot of other sets to this puzzle. There's a lot of other pieces in here. Let me just tell this thing proceed. And I'm going to hit enhance. And it's like when you're in a movie and it's like enhance and the guy's like glasses go bananas. It's going to give me something really good. So that looks fine. Cool. Keep going. So what I did was now I have the way that taskmaster works which is I like to have a set of structure for the tasks and I have my way of doing things which is more focused on the PRD and making these big docs of very detailed code snippets, relative paths, all the things. But I'm realizing that the more detailed I get, the harder it is for these things to basically just finish a work stream. Meaning signed, sealed, delivered, maybe tagged as a release, maybe document updated, maybe tests, all these things that you want to have from like first line of code done to it's out in the world and tagged and Docs and knowledge bases updated. So, I didn't want to have it be this giant thing. So, thisai folder is where all the magic happens, right? This is like NTV grips. Still going. Does this plan align with your expectations? Oof flow. Yeah, actually. Yeah, that's great. Thanks. Enhance. Enhance. I need to have a soundboard for that. But in here you can see I have tasks. I have backlog doing done. So this way that I've been working is that looks great. I have the name of the sprint or the epic the collection of sprints. And then within there I have the PRD. So, this would be an example where I was like, h I kind of want to break this up because by the time I get through the 24,000 tokens, I don't know. I want some dopamine. I want this thing to be moving. I want it shipped. I don't want to I want smaller bite-sized pieces because all these white things represent when you have a to-do. But then there's all these pieces of relevant information, environment variables, literally everything, right? So, I went a little smaller. It also has mermaid diagrams, flex. And so then I say, you know what, we need to break this up and break it into sprints, no more than 10 tasks. And it did that, right? So at first I wasn't naming them right, but then I did. So it then broke them into all these different sprints. So you can go back in time and you can see now these are 2,800 and it has all the there it is. This is the taskmaster object. So now we know what the data signature looks like. Does signature work for that? Function signature data object footprint structure carbon. My brain's breaking. It's late. But yeah, all this stuff is being here. And then we actually just get to work and we say, "All right, go do this thing." Now, once you have all these sprints in here, I'm just moving through them, right? So, I got to this fifth one and I realized, hang on, wait a second, we need to defer this because I thought of something that is even better and it needs to happen. And so, then I added something. I just tagged it as deferred, whatever. Now, this has been working great for me. And it also because this has no MCPS. I have a severe allergy to MCPs unless they run on Docker. But because this has a direct integration with GitHub, then I've just been making it have milestones. And so that would be for more so for other people, right? Because task management within your ID is fine if it's just you, but I'm going to have other people working on this. those other people are agents that are working on different branches. So, I figured one, it would be nice if a real person did jump in on one of these, but then also for educational purposes for others to be like, "Oh, yeah, having a record of this." See, the one that I do failed. No, it's cuz I haven't fixed my CI/CD. But then all this is in here and it just you can see in the last 10 hours that these were all just done by augment or whatever cursor whatever agent you're going to use. But this pattern is really nice and then it has just an absurd amount of changes and I'm loving it. One last thing because I think we're dropping frames is on the actual prompt to have it write tests. So I've found it really helpful for it to do tests and then go write the code to make the test pass. And so what I did was I just explained what I wanted and then I just prompted around and came up with something that I find helpful. write the test first and then implement the code to pass them. And then I told it we want to use bun, we want to use this for Python, all of that. And it's just been great. And then on the tool side, what I found is you need to make sure that whatever terminal you're using is selected. So it says ZH is partially supported. I haven't had any issues with it, so that's great. But that's what I put in there. And then my next step will Oh, Superbase obviously is really nice to have in there because it just can I don't know do everything which is great. But yeah, I haven't had really any issues like this is today and so 1 2 3 6 9 12 13. I've had 13 threads and it's completed seven sprints. That's on average probably 70 different tasks and 13 into 600 13 / 600 2%. So what's $50 times 002? So it cost me $1 in augment credits to do this. And the way that I came up with the work originally of having it be domain based where in our web project you can see that I have the API and all the different areas of what the endpoints are. So we have admin stuff for moderation. We have announcements. We have o with discord. We have events. We have event ids. We have registering them for the events. So when we do talks we have a feed that has comments and posts and likes. We have a GitHub O that it's doing right now that'll allow members to connect their GitHub and then people can know when things are happening. So it's like a miniature feed of that. We have the health route. We have a learning route that basically will have these paths that you can follow for the typical things that you'd want to learn in order to code. They have progress, recommendations, topics. We have the members directory with their stats. We have, and you notice there's tests in most of these. Pretty cool. We have the projects. So, if people were collaborating on different projects that they wanted to have information, so they could share them because we have share your work in Discord. But I think this would be really cool because not only could you share with internal members, but then we're going to have a page of all the members projects externally, which would mean free marketing eventually when we start to do PSO and all those things. Then we have prompts ID. How many lights do you have on those? Now, what I find interesting is because GitHub now has this models thing, which you can check out that video on my main channel. I can hit their SDK or API, I don't know. I think it's API. Yeah, I can hit their API and do read only for prompts from authorized repos. So, if I have ones that are doing really well on their evals and I want to share them, then I can put them on my profile. And then we have Stripe, we have Telemetry JK. I removed that for Sentry and I also self-hosted Sentry. That was stupid. And then we have users. So users have avatars. They have AI usage because they can bring their own key. Kind of overkill now that I'm seeing all the stuff with GitHub. But we still are going to have like our version of it. So you could just put your key in there and then test out what the prompt would do. We have learning goals. We have notification system, onboarding, profile skills, all this stuff. So, I'm pretty excited about it. This new flow, we'll have the news feed. So, all the API support basically started the day with 35 API routes and essentially all like the functionality within those total that needed to get done and it stamped these out and it's all tests are passing. So, the next step will be really just like wiring it up. And later on, what will be cool is I'll probably add TRPC to allow for a mobile app, but that's much much later. And how I got to this was just looking at a project called Midday. And I prompted around in midday with augment because it's really good at context. And so I wanted to find a project that I thought I jived with and I thought that LLMs would also drive with. So I was just asking questions. I was like, "How does this project do formatting? How does it do linting? And then another one I asked was how does it handle types? And then instead of fiddling around in a monor repo, like I've spent hours on hours when I was first doing monor repos trying to figure this stuff out and it was like going to make me cry. But then it gave me this and then I just output I write this to a markdown file like this and then I put that in my project and I say this is going to do it and then it oneshots it which is really cool. Other things that I learned when I wanted to set up this setup. When I wanted to set up this setup. Dude, I need to not record so late. Chats, threads. What was the other one? Yeah, just basically the type stuff. What else? Data fetching. Yeah, I did a ton of questions like what do all these packages do? So I get a very detailed directory reference and I say I want to do the same sort of write up for each folder in the architecture. This is yesterday. And so it gives me the full output of how theirs is set up. I say dang that's cool. And then I give it a prompt to basically tell me give me everything in a table of what all these things do. And then I bring this into my project and that sits in my folder and it's frontend backend API audit. That's where I started the day. Then I said how do I want to structure this? And that's when I got the architecture doc. And so if I pop that babson open, you can see that. So it's just the other thing. What are we doing now? Phase one. Next phase two repository management. Yeah, that seems good. Now update our sprint task list so the next dev can continue. Cool. Because I see this little long thread thing. I don't want to risk it. It is fine for one or two more, but I don't know. It gets pretty bad. So rather just not have that. Cool. That's it for this video. Just kidding. People ask questions and I answer them. studio dot. Yeah. Okay. Let's do some questions and answers. Questions and answers. Also, when I'm using augment is I just keep this little thing right here and I just keep doing what I'm doing. So, in this case, I need to go to community. Oh, this is a video I did for members. That's probably what I'm going to do in the future is I'm going to tie Gosh, now I got to write it. I need to write out this so I don't forget backlog. I need to create one. Tai YouTube membership.m MD. And I'll remember what that is. It's basically you can do memberships on YouTube. And so I can put a little lock over this and then those that are members of VI just see it on normal YouTube and that'll be cool. Oh, 15 people watched it. Nice. Let's go into questions. This is on the second channel. He says, "Turn up. Another great video. Thanks. Where can I learn more about your folder structure in today's daily update?" Agreed. Augments context engine is uncontested. Yeah, it is. Has been the only thing to provide value working these huge enterprise code bases. Yes, Cheese Sticks, you're jacked as hell, bro. What is going on? Oh my gosh. Kid is dialed. Hey, Jake. What's up, dude? Everyone go follow Florida tomorrow. I found augment to be a waste of money. I switched to the $100 D. Didn't want to worry about the price. Using cloud code. Cloud code. Cloud code. Yeah, cloud code's probably good. I don't know. Just Yeah, whatever works best for you. Exactly. It's got limits. Blackbox. What the heck is blackbox? I guarantee this guy works for blackbox. This This doesn't No, I can't. No, bro. Don't talk too much about Augment. If it increases its popularity, they'll take away my precious 14 trial for my 377. Bro, you are wilding out. When did I wilding out? I've lost my marbles. Lost my marbles. That's a How do you do that? I also wanted to Oh my gosh. I'm not teaching you. Were someone to join your community, would they have access to all the resources or the resources gated? Duh. Duh. They're the gate is joining the network. You mentioned a repo which I searched for. Okay, I'll link that after. I tried it and I liked it. Thank you. Thank you. Thank you. Thank you. Thank Okay, cool. That's on there. And then let's go over to the other channel. We're about to hit 2K. Good job. See what's so funny is I thought this video was going to go just bananas, but guess what? I'm not. There's like literally not even one banana. Instead, no. Funny how that works. Turbo starter kit. I never did that. What's with the music? I didn't like it. Okay, Dan. Dan, I thought I really thought that the music was good. Why would you make a jingle with a music editing app? Cuz I'm a freaking musician, dude. I'm literally a Vichy. Okay. Can't believe him. Okay, that's it for this video. Thanks so much for watching. Peace. Oh, like the video. And I got to do a thumbnail. Thumbnail.