Vibe Coding with Abe 💻

Hi there.

Abe Burnett, founder of Flowstate.

I thought I would do a brief video.

Well, I don't know how brief it'll be, to be honest.

My videos don't tend to be short.

But, I am working on my current vibe coding project called inkforge.

It is an AI first writing app for your phone.

Essentially.

it solves the problem of like, we all carry around these computers in our pocket, but they can be not as useful as they could be for like, things like writing fiction.

And so I have an idea for how to use AI to make writing on your phone not just possible, but fun and productive.

I call it vibrating.

so this is, it's a Flutter app cross platform so I can target both iOS and Android.

And Android.

and so I'm just working through the process of developing this app.

So I think what I want to show you today is that I'm working in the middle window here.

This is a Claude Code UI tab.

I generally prefer this perspective over the CLI experience, but, you know, both are, are equally effective.

so a lot of times I work in quad code.

Just you know a little bit about me and primary cloud code with codecs.

Sometimes GitHub copilot, I find that I get a ton of value out of the Claude code.

What is it called?

Max.

It's like the hundred dollar a month plan.

I, wish that OpenAI actually had like a $100 Codex plan because I think that that would be ideal.

because Claude code is actually like the.

Where the bulk of my usage is.

but you will see that what just happened was the one thing I've noticed with Claude code is that it will go from like 20 to 25% context supposedly remaining, and then it will go to zero in a heartbeat.

So flow state does traverse the compaction, the Claude code compaction pretty well.

much better.

Like, I mean, it does Given all the work I put into it, it does okay.

And GitHub Copilot, but it actually, it actually works well enough clot with Claude code that I don't actually find myself needing to close out sessions very often and make sure that things haven't drifted.

I have to do that more often in GitHub Copilot just because, GitHub Copilot can be a little bit of a bonehead.

you know, if you've watched my other videos, you know, I can.

I, I, have my opinions about the different tools.

and I think basically you kind of get what you pay for because with Copilot, it's way, way cheaper.

But and so I mean that's trade offs, right?

It's so cheap like $10 a month to be able to in theory access some of the world's best models.

It's not a bad deal.

And I've made flow state and with flow state it's actually tolerable.

So I'll leave it there.

but yeah, you can see basically that it crossed the compaction gap and it is actually still doing the closeout that I told it to do.

So earlier on I said, hey, I can see the compaction coming.

I want to make sure that we capture all the work that we've done.

so it finished creating like basically an avatar's bucket and Supabase that's doing an update.

Where did I do the close out?

It's maybe not showing it, but yeah, it's basically doing all the updates though here.

Updating tasks, updating session, showing what it's doing, updating to dos.

it's doing git, git commits and tagging.

And so it very graciously kind of traversed that, what do you call it, the compaction cycle.

So we're basically back up to speed.

I don't even have to like really close out and start a new thread.

I mean this, we can keep going.

and so let's do that.

Let's just do a little bit of, a little bit of vibe coding together.

Right?

So read through here.

Here's all the stuff that was completed, what was delivered, where the logs are, infrastructure complete.

So we have a bucket, we have RLS policies, we have a migration applied, we have a test script.

Okay, great.

okay, and we're ready to go on into T04.0.13 which is end to end backend integration test.

So basically iOS simulator test.

So we use clerk to sign up pen name stored profile picture uploaded to Supabase storage scene created Supabase query validates sign out sign in metadata persisted multi multi device test.

IOS+ Android simulator show same data.

interesting.

I don't have the Android simulator yet.

Session expiration test.

Let's do Perplexity.

Android for VS code flutter app.

Okay, I will have to do that separately.

I really like Perplexity.

If you haven't heard of it, for anything I would have used to like Google search for generally I start with Perplexity and then if I'm not getting satisfaction from that then I will just do a Google search.

All right, so let's go back Here.

And let's do that.

Let's say yeah man.

We're ready.

Please proceed.

T04 0.13.

So this is what we do when we're vibe coding.

It's just kind of hang around and wait for things to happen, right?

One of the challenges with clerk, which is a, a platform that I'm pretty impressed with, it does kind of integrated billing entitlements.

So you can have a variety of different kind of like models for and gates for like what people can subscribe to or buy.

I like that it's cross platform.

It has a flutter beta SDK or a beta flutter SDK.

because of that there are occasionally breaking changes.

but it's cross platform, right.

So you can kind of use 1auth/billing and kind of entitlements platform which I really like.

It's just.

It makes it way more Because I don't really want to handle say like all the entitlements and billing separately and I just like the clerk platform.

Like it's.

It's well kind of conceived.

Now what is it doing?

Remove all clerk API calls.

Why.

It is working or it was.

Thank you.

You're absolutely right.

My bad.

I'm astonished with like how fast context gets consumed with Claude code.

I think it's just that it's really, it's verbo verbose.

It's way more verbose than codecs.

I think that that is why it goes through context faster.

But it doesn't really super matter because with flow state the compaction cycle is not that scary.

What it should do is like restart the phone and boot Inkforge.

So I'm not sure why it is struggling.

Here we go.

It's trying some new things.

There we go.

It's nice.

And confirm this.

With the test.

It.

Turn off.

It looks good.

Go back to simulator.

I'm not sure how interesting it is to watch other people vibe coding, but I am making these videos just in hopes that they help someone.

Okie dokie.

Back to the simulator.

There we go.

Come back up here.

There it is.

Okay.

We are almost done with this, with this task.

It's pretty cool.

A lot of times when I'm working honestly like I will be sometimes like when I'm not recording.

I will be working in multiple windows.

Like I will be working codecs and simultaneously in Claude.

Like if I can figure out separate things to do.

Like I actually have like another project.

I am working on Ink Forge with flow state, but I'm also developing flow state, so I actually be flipping back and forth while Ink Forge is running something.

Excuse me, I'll be working on flow state, for example, or I might be working on my personal stuff and my personal laptop here while waiting for tasks to finish on my work computer.

Right.

So my work computer is doing its own things with AI and that are unrelated to all this.

but it's kind of nice to have the ability to kind of like, I hate to use the word multitasking because I know that there's no true, no such thing.

But to have something else to do, while you're kind of waiting for things to proceed.

Right.

So you can see we're about to finish this task.

it is not fast per se, but we have made material progress, quite consistently.

Right.

Like if you do this for hours a day, you're going to make a lot of progress.

Probably.

I imagine you're still probably going multiple times faster than someone like, just coding, like a human coding could do.

So like a human coding, could they do what we just did in half an hour?

I doubt it.

I mean, maybe if they're going like super fast and knew exactly what they're doing at all times, maybe, but I kind of doubt it.

So the fact that I could kind of like, you know, just meticulously or methodically kind of proceed through this, work with the help of AI, that's pretty amazing.

yeah.

And then if you do this, like I said, for a few hours a day, you will, you will create some pretty astonishing things.

Right.

because this is actually like I've been doing this for like two hours probably and this is just one of the things I've, I've completed.

So I'm making just steady progress toward having this thing up and running.

I mean it already.

Excuse me, it's only 1:00am Abe, what are you doing?

yeah, it is already running, but I'm trying to build it in such a way that the, even the mvp, the minimum viable product, is stable and functional enough that the beta testers I give it to can actually enjoy the experience and not run into too many problems that are just a result of too much bootstrapping and being too, hacky at the beginning.

Right.

I'd rather put A little extra work into making it a bit better so that it's more kind of designed with the end in mind I guess you would say.

Okay, updated all the things, tasks, good to do's.

Good.

Epic 0 100% complete.

So this is one of the aspects of it being methodical, right?

Is making sure that it keeps, it does its best, does its best to keep all the kind of orchestration documents up to date to help you stay kind of moving forward methodically.

Also the one other thing I'll mention is that I'm looking at doing working with a new swarm coding platform where you will work with multiple A.I.

agents to do coding.

And I'm thinking of ways to incorporate that plus flow state so that potentially that will alleviate you will get more speed because each agent using flow state will be more methodical.

But then if you have multiple agents kind of like working together, right.

I think that that's going to give you kind of that combination of quality and speed.

Because if you just have like each agent is methodically kind of plugging along but together I think that you will see kind of like just this massive kind of productive forward crawl.

Doing the git stuff.

Awesome.

So did our bunch of our.

We successfully did a bunch of integration testing.

things are looking good so now.

And I think it's funny, it's estimating 12 hours.

in all likelihood that will take like an hour or two of actual time.

But I think this is probably a fairly reliable estimate for what it would take a human to do.

Okay.

Alternative T0.3.8.1 Migrate Database Service to Supabase service.

I think that sounds important.

Ultrathink just means think ultra hard.

Okay, but I'm off to the races working on this new what you call this feature.

but I think we're gonna leave it there.

So this has been a 37 minute video.

Thanks for hanging out with me.

I hope that you have kind of picked up some ideas and if nothing else just enjoyed spending some time with me.

All right, well, until next time, we'll see you later.

Bye.