Troubleshooting FlowState AI 🛠️

FOUNDER OF FLOW State and this video is about what to do when Flow State doesn't behave the way that it should or the way that it usually does.

And so this should get better with some of the new Updates in version 2.4 for data science and software engineering editions and 2.5 for the unified Edition.

Excuse me, but in the, in the event that it doesn't or that, you know, it is AI, so eventually it will drift, it will have idiosyncrasies, it'll do its own thing.

Right.

So what we're seeing, what you're looking at here is actually my project I'm currently working on that I'm 5 coding.

and this is called Inkforge and it's a mobile AI first writing app.

and so what I'm going to do, because over here, I mean you can see all these other windows but we're going to ignore them for the time being, kind of make them smaller maybe and focus on this window.

Because this came to my attention because one of my customers mentioned that sometimes it just kind of hangs like, it's just like it finishes the work and then it's like kind of throws up its hands, it just kind of stops and you have to kind of nudge it to keep it going.

So like right here it's like ready for next steps.

What would you like to work on?

And it's just like, okay, so what I've done is I've added some functionality called in the troubleshooting file.

So we're just going to open that up and essentially what you want to do is start with a prompt here at the top.

So we're going to do that because what this will do just to kind of preview it for you is it's going to kind of basically read the file and then it will have Flow State kind of check in on its status, look to make sure all the hooks are in place.

I see there's an issue here actually that I need to fix.

I think that's actually what we have.

I will fix that before I send this out, before I send out the release.

Okay.

But so basically you can see at the top it's going to make sure it can access these files, that they exist.

it's also going to look to make sure the hooks that we use to integrate in with all these different AI tools that they are there and they work.

and then so we'll just see what happens here.

So let's grab this prompt here.

I'm Going to use Sonnet 4.5 and since this is a case where it's like flow state was working fine and other than just kind of like dropping the ball and like not suggesting the next thing, like I basically it's working fine.

So this is kind of like a little bit of a, a contrived example, I guess you could say so.

But you can see here that like with the upgrade to 2.4 it is doing a better job of asking questions.

Right?

Like just giving you the next thing to work on.

So I'm actually going to say resume flow state because I think in general that's what you're, you're going to want to do if it ever ends.

If it, if it does end up kind of just like stopping, because resume flow state kind of like it forces it to kind of check back in, check the different files and be like okay, what's next?

Right, so that's what we should see here.

And this is great.

This is perfect.

except for that's It can be really hard to get AI to do it exactly the way you want it all the time because it's basically saying like, you've got three priorities.

Which one do you want to work on?

And so it is kind of up to me as the user.

Right.

Like because these things are all fairly equal, in, in.

In priority from the AI's perspective.

And so I think that's what that's essentially kind of like the bottom line is that there will be cases like this when it will.

If it's ambivalent, like if it sees they're all equally important, like it's making some judgment calls.

And so in this case I think the reason why we're seeing that is just that it's like, well these things are all about the same to me.

and it does say kind of like start here.

So it is showing a little opinionation.

Yeah.

But so I could say let's do T037 next.

This is so fascinating because I've run into this like a lot with Copilot is it does a lot of this compaction which is what it basically is when it says summarizing conversation history.

It does it a lot more than Claude code or, or Codex.

it does it so much I may not continue to use Copilot.

I've almost like I've only been using it for a few days and I've already blown 14% of my quota anyway.

But I mean it is cheap.

But if it Just doesn't work very well.

Then it would be better probably to get A you know, $20 a month Claude code, or and like a $20 a month Codex, subscription.

I actually have the 200amonth Codex subscription and the hundred dollar month Claude code subscription.

And I've yet to run into any limits on either one of those.

They should be unlimited.

Well not the maximum, not the Claude code one.

But I actually have considered dropping the Codex back to the $20 a month and just using the Claude code.

$100 a month.

because sometimes I call in Codex to save my butt when Claude code gets things wrong.

but usually I prefer to work primarily with Claude code and then fall back to Codex.

Codex is okay, but like it's really good for fixing things I think once it sees kind of the idea.

And Sonnet is really Good at Sonnet 4.5 is really good at kind of laying the groundwork I think.

So I don't know.

Just some thoughts, random thoughts from me on AI stuff here and strictly speaking I don't have to show you this, but why not?

This is because session is, was historically not being kept up to date as much as tasks.

and in fact like even with the new additions sessions will be a little bit like it will trail tasks just because updating sessions all the time gets really slow.

Like miserably slow.

So basically the sweet spot I found was like that methodical is where tasks stays updated very frequently and sessions gets updated every like when something major gets completed.

All right.

Yeah, that's basically it.

So that is troubleshooting.

So and if you want to know more about troubleshooting, what else you can do.

so there are various AI instructions.

Right?

so you can see that.

but you can read through this, you know, like a lot of isn't.

A lot of it is intended for the AI and essentially the reason I put the beginning there or the prompt at the very beginning and did not include like a whole bunch more user stuff.

It's just because I put the burden of helping you on the AI.

So the idea is that if it runs into anything, if it discovers any gaps or problems, then it will help you solve the problem instead of asking you to read a lot of documentation and do a lot of manual work.

So that's the idea.

let me get another thing I need to check and fix but I will fix these things before I push out the new additions.

Cool, thanks for watching and let me know if you have any issues with this, you know, room.

There's always room for improvement.

And I'm happy to, you know, keep on.

Keep on refining things.