October 9, 2025 • 43 mins
In this episode of JavaScript Jabber, I sit down with AWS’s Clare Liguori and Erik Hanchett to talk about Kiro, a brand-new AI-powered IDE that’s reimagining the way developers build software. We dive into how Kiro takes “AI-assisted coding” to a new level through spec-driven development — a process that focuses on defining requirements and collaborating with AI to break projects into clear, manageable tasks.
We unpack what sets Kiro apart from tools like Cursor and Copilot, explore its supervised vs. autopilot coding modes, and even talk about how it handles UI design, planning, and complex legacy codebases. Clare and Erik share behind-the-scenes insights on how Kiro was built using Kiro itself, what’s coming next for the platform, and how developers can join the early-access community to help shape its future.
🔗 Links & Resources:
Become a supporter of this podcast: https://www.spreaker.com/podcast/javascript-jabber--6102064/support.
We unpack what sets Kiro apart from tools like Cursor and Copilot, explore its supervised vs. autopilot coding modes, and even talk about how it handles UI design, planning, and complex legacy codebases. Clare and Erik share behind-the-scenes insights on how Kiro was built using Kiro itself, what’s coming next for the platform, and how developers can join the early-access community to help shape its future.
🔗 Links & Resources:
Become a supporter of this podcast: https://www.spreaker.com/podcast/javascript-jabber--6102064/support.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:05):
Hey, folks, welcome back to another episode of JavaScript Jabber.
This week, on our panel, we have Steve Edwards Yo yo.
Speaker 2 (00:13):
Yo, coming at you aj style from just outside Washington, DC.
Speaker 1 (00:19):
Yeah. I was trying to come up with your name
and I was looking at Eric, so my brain kind
of went wait, wait, wait wait. I'm Charles max Wood
from top Endevs. And this week we have two guests.
We have Eric Hanschett, Hello, Hello, welcome back, and we
also have Claire. I didn't get your last name, Claire
well gory Oh cool, sounds Italian, so you're probably awesome, all.
Speaker 2 (00:43):
Right, So I sounded French, but that's just me.
Speaker 1 (00:48):
It could be anyway. Do you guys want to just
introduce yourselves real quickly? Got Eric on before, but Claire,
you're new to us, so we want to know how
awesome we are.
Speaker 3 (00:56):
Sure. I'm Claire. I'm a senior Principal engineer at as
in agentic AI, and I work on all seeing developers
and agents basically hero ID and the strands Open Source
Agent SDK.
Speaker 1 (01:14):
Cool.
Speaker 4 (01:16):
Well, and I'm Erica. You've been on here a few times,
a senior developer advocate at AWUS Huge VIEWJS fan it's
probably the last few times you may have heard me
on this podcast talking about it and all the work
I've done in that community and in the JavaScript community.
So glad to be here today and talk about some
really exciting that we released recently. That Claire can talk
(01:38):
a lot about Kiro, which is our new AI I D.
Speaker 2 (01:42):
So a little bit of background before we get started.
So we had Eric on a month and a half
ago somewhere around there. I forgot and found out later
that he had been planning to talk about Kiro, but
it turned out it wasn't quite ready for release at
that time, so we sort of scrambled and talked about
other things sounded we sounded really good thought. But now
Kiro's has been released, and I know I've heard a
(02:04):
lot about it on other podcasts, you know, in other
blogs and so on, and people have been really raving
about it. So we're finally here to talk about it
officially and all it can do for us.
Speaker 1 (02:16):
So I'm curious. Sometimes it's obvious why people name stuff
what they name it. Right, It's like this does email,
so we called it email thing? Right? How did you
come up with kiro?
Speaker 3 (02:28):
That was a AI assisted process? Actually? It actually means
crossroads in Japanese, and so we were looking around for
a word that would reflect the sort of new development
style that we have now. Our development has changed so much,
so we were at a crossroads.
Speaker 1 (02:52):
Gotcha. It's funny because I go for walks and I
talked to Grock the whole time, so I know how
this goes. Sometimes it's like, oh, that's a good idea,
So do you want to explain what it is and
how it works, maybe how it's different from some other
things that it sounds like do the same thing.
Speaker 3 (03:08):
Well, Kiro is an AI powered ide like Eric said,
but one of the things that makes it different is
that we've all been kind of vibe coding for a
bit now, and one of the things that we were
noticing is that people really struggled to get what they
wanted out of vibe coding, especially when it came to
(03:29):
more complex tasks. And there was a kind of an
aha moment for me when we and was the reason
that we started Kiro, which was what if developers instead
of saying kind of commands to an AI assistant of
how to change their code, what if they just declared
(03:49):
what they wanted? What if they defined requirements and then
worked with the AI on requirements and how that broke
down into individual tasks. And that was really the birth
of what we call spectrum and development, which we have
dedicated UX around in KIRO. That's really what makes it different.
Spec based development is kind of what it sounds like.
(04:11):
You create a specification for what you want. That could
be a brand new application, it could be a feature,
it could be a massive refactor in your code base.
But you go through this process where you define the requirements,
the use cases, user stories. You can have the AI
generate all of those, because I remember writing these by
(04:31):
hand ten fifteen years ago and it wasn't that much fun.
And then you go through a process of working through
the technical design with the AI and finally what tasks
that breaks down to, and then you have a lot
of control over steering the actual implementation of those individual tasks.
And we've heard from folks that this has enabled them
(04:53):
to at the end of the day, produce much better
quality code and actually get to what they wanted. I
said of this very frustrating experience with vicuding, we were
hearing about where it's like you go ten cycles and
that it's almost like the AI is not getting it
getting what you want.
Speaker 1 (05:12):
Yeah, I've experienced that a little bit. I've been using
cursor and a co pilot. I'm depending on what i'm
working on. And you talked about it breaking it down
into lists, And it's funny because cursor will give me
a these are the things I'm going to do, and
then it'll check them off as it goes right. And yeah,
it doesn't always get it right. If it's something simple,
probably the vast majority of the time it gets that right.
(05:34):
But yeah, the second it gets complicated, it's like, Okay,
this isn't working. Sometimes I'm still not quite sure what
it did or how to fix it, and so I'll
ask it to fix it, and then it still doesn't work. Yeah,
So I get that. It sounds like this is a
little bit more in depth than sort of the process of,
you know, giving it a single prompt and then having
(05:55):
it come up with I'm going to do this, and
then I'm going to do that, and then I'm going
to do that and then I'll be done.
Speaker 3 (06:01):
Yeah, and it still starts with a single prompt. But
one of the things that I've noticed is that it's
really hard for us to describe what's in our head
from the get go, from a blank prompt text input box.
And so often what I find is that there's this
iterative process of effectively the AI trying to guess what
I meant by this simple sentence and then being able
(06:25):
to iterate on that and say, no, no, that's not
exactly what I meant. I actually meant this, or oh,
that's an interesting idea I didn't even think of. And
so that process of creating this spec is really helpful
for refining what it is you want. Often if you
go to Vibe coding, it's kind of like you know,
when you have an engineer who's a bit a bit fresh,
(06:48):
and they always jump right to writing the code before
actually thinking about the problem. That's kind of the Vibe
coding experience is the AI jumped into writing the code
before kind of understand the problem. And so this spec
driven development helps you to kind of make sure you're
on the same page and that you know everything that's
(07:08):
in your head of what you're thinking about the application
or the future is actually out on paper without necessarily
having to type it out all yourself. You kind of
just go through this iterative process of the AI kind
of guesses writes the spec for you, and then you say, no, no,
that's not exactly what I meant.
Speaker 1 (07:26):
Yeah. It sounds a little bit like the process I
went through more as a freelancer than really is a
full time employee, where you know, I'd have the client
come to me and say, hey, we want this feature,
and then I'd ask questions, do you want it to
do this? Do you want it to do that? And
you want it to look like this? Do you want
it to blah blah blah, right, and then there was
further iteration once we got going where it was okay,
(07:46):
I have a rough version of this, and then they
go look at it. Oh, no, this isn't quite what
I wanted. And so by the time we're done, yeah,
I had a fully fleshed out understanding of what they
wanted and could build it. And it sounds like that's
what this is doing.
Speaker 3 (07:58):
Yeah, And I love that idea of the rough version
of it, is this what you wanted? Because so often
we don't know what we want until we actually see
it when it comes to application development. And you know,
we've seen across moving from Waterfall, where we tried to
you know, have all the requirements ironed out at the beginning,
(08:19):
and that never really worked because those were not Those
were never the actual requirements that you ended up with
at the end once you saw the application. We went
to agile, where we're attempting to move into sprints, where
we have some capability of refining those requirements as we build.
But even with something like agile development, it still takes
(08:42):
so long to get to something that's working that you
can actually show to a customer, show to a user.
So it's going to take multiple months. And one of
the powerful things about AI assisted coding is you can
get to that tangible result so much faster. You can
build entire application may not work exactly right, it may
(09:02):
not be exactly what you had pictured in your head
to begin with, but you can actually play with it.
You can you can touch it, use it, and go
back to your requirements and say, oh, actually this this
part wasn't exactly right, or oh we forgot this, let's
change these a little bit and regenerate. And so that's
been really powerful for getting into kind of prototype driven development,
(09:26):
and I've found that that ends up with a better
product at the end of the day because you are
able to really really use the product very early on
while you're still forming the requirements.
Speaker 1 (09:37):
Really yeah, I'm kind of imagining how this flows. I'm
a little curious once it has the spec if it's
something that's a little more complicated. One of the things
that I've experienced more with Cursor is that it will
go and it will do the whole job right, and
so it gets done and I'm like, I have a
(09:59):
lot of stuff.
Speaker 3 (10:00):
Yeah, yeah, yeah.
Speaker 1 (10:02):
Does it do that or does it kind of so
march through a piece at a time there too.
Speaker 3 (10:08):
We call that supervised versus autopilot mode. So sometimes you know,
you just want to get something done right. You just
want a really simple tool and it doesn't need to
be exactly right. And I find that autopilot mode is
going to be the fastest way. That means exactly what
you just said. It's just going to plow through the work.
(10:29):
It's going to completely build the entire thing. Supervised mode
is where you want to give feedback one step at
a time. So inspec mode, what you can do is
it generates a task liss similar to what you were
saying that the AI says, here's what I'm going to do,
but it will actually stop for your feedback after each
(10:52):
individual task, and so you can say, whoa, whoa, you
are starting off on the wrong foot here, let's rewind
and get this right. And I find that that does
end up with more of the code that I already
wanted because I'm steering it along the way. In between
these little individual tests. I will say that I've noticed
(11:19):
AI models really like to do a lot of work.
And so what I find is when they when they
generate these spec files, it's following all the software engineering
best practices. It's got high unit test coverage, it's got
integration tests. It might sometimes say I'm going to try
to deploy this before before deciding I'm done. And that's
(11:44):
really great if you are building a production app. Sometimes
I'm just building a little side project or I'm building
a little utility that I just kind of need to
get done and I don't need integration tests. So being
able to review that list of tasks for it goes
off and does a lot more work than I wanted
it to do is also really helpful.
Speaker 4 (12:07):
Nice I feel like it's like an eager intern. It
just wants, it wants to help you so much and
do everything for you, and you can definitely be like, no,
let's supervise this one at a time. YEA, let's list
delete things we don't need. We don't need three thousand
integration tests.
Speaker 1 (12:25):
Definitely. I love the eager intern analogy. You sent me
for coffee and I came back with doughnuts and sandwiches.
Speaker 3 (12:32):
And the whole coffee machine did.
Speaker 1 (12:35):
Your dry clean. So so do you use it now
or go ahead?
Speaker 2 (12:43):
I was going to say so in reading through the
instruction chareing what you were just talking about. It sounds
like there's a lot of it's like dock writing, almost
like you're planning and figuring out I'm going to do this.
How much of it? How much does it actually write
code versus just doing the planning? Is it all of
the above? I mean could say, for instance, if you
are a design challenge back in engineer like me or
(13:06):
a code writer like me who has no design chops whatsoever,
you know, and I actually have a project I'm working
on that I really would love to do, this can
do something like come up with at least a somewhat
workable UI design based on what you're doing, you know,
in terms of page layouts and elements, and then you
naw bar structured tables formed that kind of stuff as well.
Speaker 3 (13:28):
Yeah, it can both write the specs and write the
actual code for you. And the technical design step is
really helpful for some of the UX as well, and
kind of the code architecture, so you can say, what
front end framework do you want to use, what kind
of component library do you want to use? You know,
compared to I mean when I first started out in
(13:52):
the industry, I was literally writing waterfall requirements documents. What
I like is that it's not so much about you
writing these documents as they are kind of artifacts for
a shared understanding between you and the and the AI,
and so you don't have to actually write any of it.
(14:13):
You can keep prompting that, you can keep prompting in
KIRO until you have the requirements that you want, or
you can you know, kind of edit them yourself where
need be. But it's really helpful and really the whole
life cycle of kind of ideation and then design and
then actually writing the code.
Speaker 1 (14:33):
Yeah, I'll just jump in. I assumed that I guess
I've been working from the assumption that it's as capable
as any of the other systems out there, right, So,
like I said, cursor and copilots seem to be able
to do the UI stuff just fine, because I have
the same problem, right, I work almost exclusively on the
back end. In fact, for my full time job, I'm
basically building APIs and then somebody else is doing all
(14:55):
of the front end work. And so when I'm working
on my own projects, yeah, I tend to let it
do the UI UX stuff and it does fine. I mean,
it's not always beautiful or whatever. But the thing is
is I can more easily than tell it I want
this over here, or hey, this isn't working quite the
(15:17):
way I expect. And a lot of times again back
to the point of you have to see it before
you know what you want. Like I'll go in and
I'll just do something instinctively and I didn't tell it
I needed that, But then I'm like, oh, I need
it so that when I click here, it does this
and so. And it's usually pretty good about that.
Speaker 3 (15:36):
And you can also give it. You can give it
kind of a Napkin diagram of what you want the
UI to look like, and it will actually translate that
into requirements or design statements. So that's pretty cool as well.
Speaker 1 (15:51):
So one thing that I'm wondering having used. Like I said,
I've used Cursor and I've used Copilot. I'm working on
figuring out how to get Copilot to work with Emacs.
So I like that a little more than Visual Studio code.
But Cursor's based on DS code. Is that is that
what you did with kiro or is kiro something else?
Speaker 3 (16:09):
It's based on code OSS similar to the Yeah, so
it feels very familiar. But we were also able to,
for example, with this this speck layout. One of our
hopes is to a little bit take the focus of
the environment away from just the code. So one of
(16:30):
the things that I noticed as we were working with
code OSM is it's interesting the way that the colors
are where code is very much the focus. So in
that I use a light a light mode, and all
everything in the UI is gray except for the code,
which is in white white background. And so we were
(16:54):
we were looking at this and thinking, you know, actually,
you don't want necessarily the code to be the focus
in your ID, and I think we're not We're not
quite there yet. I think our first step is with
this kind of dedicated UX around spec based development, where
the focus is on the spec the design and the tasks.
But I think my hope at least is that the
(17:17):
ide becomes a place it is not just about writing code.
It becomes a place that is more tailored around, you know,
defining what you want, and then the code is is
kind of an artifact that comes from that.
Speaker 1 (17:34):
MM I'm trying to imagine to myself what that looks like. Also,
you're the only person I know that uses a light mode.
Speaker 3 (17:47):
I just don't like staring at black all day.
Speaker 2 (17:52):
So can we maybe go at a high level? I
don't know how detailed do you want to get. Granted,
everybody can't really see the ID since we're not screen
sharing with you can just sort of walk some steps through.
Speaker 5 (18:04):
Let's say you're starting a.
Speaker 2 (18:06):
Project completely from scratch and you want to build some
you know, maybe something more complex than a two do app.
You know, who knows what are the steps? Because now,
for instance, if I come, I'm looking at the ID
right now, and it looks like you have two different
ways possible ways to start, it gives you the option
to be the vibe respect right, So one is chatting
and one is planning first, So it looks like those
(18:28):
are two different ways you can start.
Speaker 5 (18:31):
Am I reading that correctly?
Speaker 3 (18:32):
That's exactly right? Yeah.
Speaker 2 (18:35):
Okay, So let's say I want to start with the
vibe mode. So what am I doing or how what
are the steps I'm taking.
Speaker 3 (18:43):
That's vibe coding, as I think we're all kind of
familiar with. Now you're you're prompting the AI to generate code,
it's going to you know, if you do have any example,
any existing code in there, it's going to take a
look at that and bring it into the context. But
if not, it's just going to start writing code for you.
Speaker 2 (19:02):
Right, So there's this chat panel down here, so this
is where you start saying, hey, yeah, this is my code.
I want to add this feature that's right, and make
it look like this. This is my stack, you know,
and I'm using you with you know, a level in
the back end with nursia, jas and tailwind, et cetera.
That's my preferred stack, yep. And then it's going to
(19:23):
start generating what is it going to start generating documents
there that with the output, you know.
Speaker 3 (19:29):
Yeah, so in vibe mode, it's just going to start
writing code. It's that that very familiar experience I was
seeing before, where it's just going to jump right into
writing the code and then kind of iterate from there
of trying to get the AI to update the code
into what it is you want it to build. And
so I find that that's very useful. I find that
(19:50):
the so the Cure team built uses kiro to build kiro,
and I find kind of across the team that we
tend to jump into vibe mode if we know exactly
what we want. You know, we know the code that
needs to be written. We're just kind of steering the
AI to generate that code. And it's a pretty simple,
(20:12):
simple problem. But if we don't know what we want,
we don't know exactly what code needs to be written,
will jump into spec mode or if it's just a
more complex problem. And I think that that takes a
little bit of intuition today, It kind of takes a
(20:33):
little bit of playing, and it requires knowing what do
I think the AI is going to be good at
kind of one shot right vibe mode to get what
you want. You want it to be pretty close to
one shot. You don't want to have to do a
ton of prompt iterations. But if it, you know, you
kind of get this intuition of what is simple enough
(20:54):
for me to give to vibe mode and then anything
more complex than that that needs to be broken down
into individuals or they don't exactly know what I need yet.
It is a really good starting place for a spec.
Speaker 4 (21:08):
And maybe I'm jumping ahead here, but as we're using bimode,
as Claire said, there's it reads the files, but you
can't create steering files or kind of the same thing
you might see and other code editors like rules files
that you can set up so it follows certain patterns
as you're vibe coding back and forth. And you can
even have here generate these steering files for you to
(21:32):
start off with.
Speaker 3 (21:35):
Yeah. I find that that's really the first thing to
do if you're starting in an existing project is generate
the steering files because it creates this really great summary
of the code base and what it is and what
it does and what the architecture is. That really helps
you to save on contact space, so it's not having
to read a ton of files every time.
Speaker 4 (21:59):
Yeah, outputs better code.
Speaker 1 (22:02):
I feel like, Yeah, so a couple of other questions
that I have, I'm going to change the topic a
little bit. So both Copilot and Cursor allow you to
change the model you're using. I'm assuming that that's the
case here too. I can say I want to use
Claud and then okay, Claud's not cutting it for me.
(22:23):
I want to use GPT or is there some kind of
auto selection that you do?
Speaker 3 (22:28):
Currently the model that's available in the product is clouds
on it for.
Speaker 1 (22:33):
Okay, and I can't change that.
Speaker 3 (22:37):
That's right, But we currently in previews, so things change.
Speaker 1 (22:43):
Right, Yeah, and only get better, folks. And then the
other question I had was, so both those other systems
that I've used you get for what you pay. So
my employer is paying for Copilot, and so that's why
I'm using cursorund my own stuff is because get hub
didn't give me a good way to differentiate between credits
(23:06):
that I pay for and credits that they pay for.
So I've looked at setting up a separate getub account
just for their stuff. And then anyway, but I guess
my question is, so you always have a limit on
the number of credits or premium credits you can use?
Is that the same for Kiro? I looked at the pricing,
patient looks like that's more or less the case.
Speaker 3 (23:26):
But yeah, we have I guess four different tiers free
pro pro blust power and very similar to other pricing
plans you've seen in the space, you get a certain
number requests or credits that come with that, and then
over once you go beyond that for the month, you
can enable per request or for credit pricing.
Speaker 4 (23:50):
And since we are in preview, like Claire said, things
are still in flux, things might be changing. We have
been listening really closely. You have a discord community and
there's Reddit communities, lots of different communities we're looking at
hearing the feedback has some really great people in those
threads every day listening to that. So we're constantly listening
and so there may be updates, may not be updates,
(24:12):
but stay tuned on that and check our blog on
all the pricing details. Since it's in preview right now,
it's under a wait list. So one thing is we
had lots of great feedback when we first launched it.
We had actually so many people want to try it,
we put up a wait list and so we're slowly
going through that. Weitlist, adding more people on and then
(24:32):
we'll have more information about that soon.
Speaker 1 (24:36):
So one other question that I have with this, I
feel like i'm hogging the mic, but I'm going to
keep asking my questions. So I'm kind of imagining. So
the Vibe mode sounds a lot like what I've seen
other people doing with the other agent systems, right, and
so you know, I don't necessarily see a major advantage
one way or the other between this and those. But
(24:56):
then you've got that spec mode, which is something that
is different, right, And so I'm I'm trying to think
to myself, Okay, when will that really really deeply pay off?
Is it those super complex features or tasks? Are there
other things that I'm just not even thinking of where
you're going to say, oh, well, we tried this with
(25:18):
our you know, with our workload, and it made a
major difference here and maybe not so much of a
difference on some of the other stuff we did.
Speaker 3 (25:26):
I mean, in my experience, I would say I get
higher quality code at the end of spec mode. With
five mode, I find that either I have to put
some of the practices in steering rules, like you have
to read, you have to write tests, you have to
run those tests. I find that at the end of
(25:48):
the spec mode, the AI has already done that for me.
And like I was saying, sometimes that's overkill. When it
puts that in the task list, it says, you know,
I'm going to write tests and run tests at every
step of these data. And sometimes I say, no, it
just remove all the tests. But but when I'm writing
something for production, I find that I get higher quality.
I find that I don't have to iterate on the
(26:11):
code as much once it's finished, especially if I'm steering,
just steering it along the way, And so I find
I end up with I guess the right thing the
first time more often. And I would say the other
one is I'm trying to think of We have some
examples from the Cure team using Kiro to build Hero,
(26:34):
especially in UH some of the code oss UH code base.
This is a massive ten year old code base, and
there's some changes that we've had to make in the
in the in our our fork of it. It is
(26:56):
really hard to understand that. And so I was thinking about,
you know, so many developers having to work through a
massive legacy code basis, which we know there are many
of in the world, and having to trying to make
a change to one of those code bases. This is
like incredibly scary thing, right, and our experience was that
(27:17):
that process helped us to work with the AI to
build an understanding of where those changes needed to be made.
And that was because it was able to kind of
have this planning stage where it looked through a lot
of the code, summarized, a lot of the code, figured
out all the places that needed to be changed for
(27:38):
a particular feature, and that was helpful for the developer
learning as well. Right, they didn't they didn't actually know
where where all these changes needed to be made, so
it was really helpful for that, you know, both the
steering files that were generated for that code base, but
also just going through this planning phase was really helpful
for the developer doing that.
Speaker 1 (28:02):
Cool.
Speaker 2 (28:05):
Well, yeah, I mean that would make sense, right that
anything more you put into the planning, the better your
execution is going to be on the back end, while
at the same time understanding you know, I just am
working through a large project at my day to day
and no matter how much you plan, there's stuff that's
going to change as you go along. You're going to
find out a library doesn't work or doesn't have what
(28:26):
you need, or here's a better option, or you have
to totally scrap it because of other considerations. So I
guess if kiro can handle the changes, which it seems
like it can, where you can go and say, okay,
this change, now I need to rescope this and have
this then yeah that's awesome.
Speaker 1 (28:45):
Yeah. And the fact that you can steer it as
it barfs out all that code.
Speaker 2 (28:53):
Yeah, bar is that a technical term?
Speaker 3 (28:56):
Trap?
Speaker 1 (28:58):
It is when I use cursor.
Speaker 2 (29:00):
So just from a training standpoint, I've been poking around,
you know, just on YouTube. If I search, I'll see
a bunch of stuff on hero. Do you guys have
any particular training videos or step in your channel.
Speaker 5 (29:13):
On kiro on like video stuff?
Speaker 2 (29:16):
Okay, here's how you do this.
Speaker 4 (29:19):
I could jump in here in this one.
Speaker 3 (29:21):
Yep. Yeah.
Speaker 4 (29:22):
So the full developed Rabits team, We're really excited. I've
been creating a lot of content. We have our own
Twitch channel, so we've been creating weekly every other week videos.
Claire was just actually on one couple of weeks ago,
I believe, and we've been creating. We have a YouTube
channel for kiro too. We have a bunch of videos there,
(29:44):
and we have blue Sky and x account. So yeah,
we kind of hit all the channels up to try
to get out and also just going to the kiro
dot dev website and then looking at the blogs in
the docks a great way to start.
Speaker 3 (29:58):
And the discord community is so active. That's been one
of the the nicest surprises almost of this lunch. You know,
we wanted a way to connect with individual developers, right,
but I've been so excited to see really a community
(30:19):
forming and people helping each other on the Discord. So
that's been amazing to see.
Speaker 1 (30:27):
Yeah, I have to say, with most of the things
that I've worked on, when the people are great, it
makes it just that much better. Like it, it's fun
to deliver something cool and it's fun to see that
it made a difference. But when you're able to connect,
I don't know, at least for me, it's a whole
different thing. Yeah, So where does this go next? Like,
(30:51):
what are you putting into kro next? Is there going
to be a you don't even need a developer any more? Mode?
I hear people worry about that all the time.
Speaker 3 (31:00):
I mean, some things that folks have been asking us for,
especially in the Discord, We're listening very closely, things like
background tasks folks have been asking for, where maybe you
do want to set it and forget it for a
little bit and let it go for ten to fifteen minutes.
We've been finding that that mode is useful when you
(31:23):
have a project where the AI can iterate for a
long time. So, for example, we have a project internally
that is written in RUSS that was done entirely with
AI coding, and the Rust compiler gives such good feedback
about the code that you can kind of let it.
Let the AI go for ten minutes and it will
work out all the bugs because the Rust compiler is
(31:45):
giving it such such good feedback, and so you can
let it go for ten minutes and the code that
comes out might not be you know, massive amount of
code of blab of code for you to review, but
it's gone through so much testing it's definitely going to be.
It's almost always right when it comes back. But you
(32:10):
also don't necessarily want to be you know, running your
local laptops to you for one hundred percent. For that
whole time, the folks have been asking us, you know,
how can we run it? Have the task maybe run
on a different machine, like on an easy to instance
something like that, or just you know, multiple threads going
on in their local laptop. Let's see some other things
(32:32):
that folks have been asking us about.
Speaker 1 (32:34):
I'm imagining somebody using that Kiro build me Twitter.
Speaker 3 (32:38):
Oh yeah, yeah, somebody already me a YouTube clone that
they built with Kiro. But let's see what else. There's
a lot of opportunity for improvement in the spectrum and development.
I think we're just getting started there. We have some
really interesting ideas about how to prove that the code
(32:58):
that was generated actually meets the requirements that you set.
I think we've all seen that. Sometimes models will say
that they have met your requirements and they have not
met your requirements, and so we have some really interesting
ideas there.
Speaker 1 (33:13):
Yeah. I've experienced that too, where it says I did
it and I tried it and it worked and no.
Speaker 2 (33:21):
All right, so Eric, what is the Sorry? Go ahead, we're.
Speaker 1 (33:25):
Getting close to the time for picks, so when you
to start writing down, But that's fine, go ahead and
ask your question.
Speaker 2 (33:31):
So, Eric, what is the AWS channel on YouTube or
what is the channel for Kiro?
Speaker 4 (33:36):
Yeah, there's a Kiro channel on YouTube. I think it's
kiro dot dev. So if you go to YouTube dot
com slash at kiro dot dev you.
Speaker 3 (33:45):
Will o t k Dev.
Speaker 1 (33:47):
Yeah, thank you.
Speaker 4 (33:48):
Yeah, k I R O d O T DEV. That's
a good call it yep.
Speaker 1 (33:53):
Okay, And do we get the visuals on there? What's that?
Do we get to see your faces on there or
somebody else if.
Speaker 4 (34:03):
You look there today? As of this recording, I'm not
sure when this is going out, but I said today, Yeah,
I just did a video with James Ward. It's extending
tirost capabilities with MCP. We didn't even get into this,
but it does have MCP capabilities, which is I'm sure
you guys done shows on it. But yeah, we go
through in this in the latest video in the channel
on how to I add some fun MCP servers how
(34:26):
to set that up? So yeah, and there's a lot
of interviews, a lot of great information on there to
check out.
Speaker 2 (34:31):
Wow, when you're dressed up too, it looks good.
Speaker 4 (34:34):
Yeah, I got I got to wear the sports tacking occasionally,
not always just wearing the.
Speaker 1 (34:39):
Black on black.
Speaker 2 (34:40):
Look, that's really good man.
Speaker 1 (34:44):
All right, Well let's go ahead and do some picks, because,
like I said, I have a hard stop. So we'll
go ahead and start with Steve. Steve, what are your picks?
All right?
Speaker 2 (34:54):
So don't have anything else outside of the dad jokes
of the week, So we'll dive right into those. So
someone threw a can of sodat maybe the other day,
but I'm okay, it was a soft drink. Simple question.
A couple of questions here. Why do hummingbirds hum because
they don't know the words.
Speaker 3 (35:16):
Thank you?
Speaker 2 (35:18):
And finally another question, why do people say tuna fish sandwich?
Nobody says chicken bird sandwich. Those are the dad jokes.
Speaker 3 (35:31):
Of the week.
Speaker 1 (35:35):
All right, you said you didn't have anything else, that's correct.
All right, I'm gonna go next. So Claire, you haven't
been on the show before, so I'll just explain. We
just shout out about whatever, and so Steve shares the
best of the dad jokes, and then you know, we'll
do other things. So I usually pick up board game
(35:56):
or a card game, and so I'm going to pick
a called The Infiltrators. Infiltrators came out in twenty thirteen
or twenty twenty three, sorry, and it's done by Pandasaurus.
They have other games. It is a two to five
player card game. And what it is is you have
a certain number of I guess spies that you're trying
(36:19):
to or mafia members that you're trying to pinpoint. And
so what they are is they're face down cards in
front of each player, and so as you look at
the there's a delay in my videos striting crazy. So
as you play, then what you can do is you
can so I can give a clue on my cards
(36:42):
or somebody else's cards, but if I play it on
somebody else's card, then I get to draw a card
back to my hand. Otherwise I have to take a
turn to get more cards eventually. And so usually you're
cluing other people's cards, but sometimes you clue your own
because you know what yours is and nobody else does,
so you know. Then you can kind of give people
the you know, instead of them guessing over and over again,
(37:04):
you can just basically tell them what it is based
on what they already know. And so if I put
put a card on somebody else's they can turn it
one way or the other that says there's nothing in
common with this card, or there's something in common with
this card, and it basically what it can have in
common is either the numbers on the card. And so
let's say that you have the fifteen that you play
(37:26):
over there. Well, the fifteen shows all of the factors
or multiples of the number, right, So if you play
a three as three, six, nine, twelve, and fifteen, if
you play the fifteen, it has one, three, five, and fifteen.
And so if my card is a one, three, five
or fifteen, I don't think there were ones anyway. If
it has any of those numbers, then I'll put it
(37:47):
down and say there's something common, or if it's the
same color as something common. But if there's nothing in common,
then you can eliminate a whole bunch of cards. You
can you say it's not blue anyway. So it's it's fun,
and what you're trying to do is you're trying to
expose what all of the cards are. And so one
of the things you can do on your turn is
instead of putting down a clue, you can say you're
(38:11):
supposed it has a little gun prop and you're supposed
to execute the card. But we always just say that
one's the three of red, and then you know, whoever
had it flips it over and says you're right, or
if you're wrong, then there are consequences for that. But anyway,
it's a fun game. Board game geek waits at at
two point one point one. That's out of five, and
(38:32):
so it's it's pretty approachable for people for the casual gamer.
Sometimes I like the deep dark games that are like
three or three point five and weight. But you know,
if you just want to pick up a game you
can play in a half hour, that's a great one.
And it's just it's a deck of cards, so anyway,
not a like a a poker deck, but it has
(38:56):
its own deck. But anyway, super fun. And so I'm
going to pick that we played it, I don't know,
like six times the last time I got together with
my buddies. We get together and we play every Wednesday.
So we're going to play tonight. We'll see what we play.
Speaker 4 (39:11):
What's the name of the game again.
Speaker 1 (39:13):
Infiltraders, Traders is spelled like traders like people who betray people,
so it's not the typical spelling of the word. And
then I'm also going to pick While I was in
Atlanta last week, I was in the hotel trying to
just relax, and so I was looking for a show
to watch and I picked up The Twisted Tale of
(39:35):
Amanda Knox and Amanda Knox is an American student who
was studying in Peruja. When I was a missionary in Italy,
I didn't live that far from Peruja actually anyway, and
that's been kind of fun for me because like half
the show's in Italian with subtitles because the Italians are
speaking Italian. But anyway, so she gets arrested for murder
(39:56):
and so they're kind of going through the trialing their
five episod So it's out as we record this and
it comes out every Tuesday or Wednesday. But yeah, really
really enjoying that, and so I'm going to pick that show.
I've also been watching The Terminal List and that has
Chris Pratt in it. That one's a little bit more violent,
but I'm enjoying that one as well. Twisted Tale of
(40:16):
Amanda Knox is on Hulu, which is included in our
Disney I've been watching on Disney Plus, and then a
Terminal List is on Amazon Prime. So yeah, I'll pick
all of those. I'm trying to think what else. But
I'm going to be speaking at commit Your Code conference
in at the end of September, and I believe this
will come out before then. So if you're going to
(40:36):
be there come find me. And then on the next
week I'm going to be at Rocky Mountain Ruby Conference
in Boulder or no, that's two weeks later, but anyway,
so I'm going to be around if you're in Atlanta.
Price Picks is based in Atlanta, so I wind up
out there. So just email me and say, hey, I'm
in Atlanta. Let me know when you're coming out next
and I'll let you know. But yeah, those are my picks, Eric,
what are your picks?
Speaker 2 (40:58):
I good?
Speaker 4 (40:59):
Well, So last night I went to the movie theater
with my wife and I guess Hamilton has a ten
year anniversary movie that you can watch, So we watched
the ten year Anniversary Hamilton movie. That included like ten
fifteen minutes the beginning of like interviews with the cast
and everything. So that was kind of a fun evening
(41:20):
to check it out. Of course, you can get Hamilton
and Disney plus the movie, but it was nice to
check that out in the movie theater, which I've never
watched Hamilton. I've watched it at in Broadway, but I
have not watched it in a movie theater, which so
that was fun. So that's my first pick, and then
I guess the second pick would be I just went
to Singapore for an AWS event a few weeks ago. Man,
(41:42):
I love Singapore. I just love the food, the hawker stations,
the all the amazing gardens. So I highly recommend people
who want to travel to a nice destination. Have I
went to Gardens by the Bay, which had this crazy
light show. They do it like once every day or
twice a day, and so yeah, great weather, great food,
(42:05):
highly worth checking it.
Speaker 1 (42:08):
Awesome, Claire, what are your picks?
Speaker 3 (42:11):
Let's see. So I just finished a book called Custodians
of Wonder by Elliott Stein. Ancient customs, profound traditions, and
the last people keeping them alive. So this guy went
and found people who are kind of the end of
the line. There's a woman in Italy that makes a
certain kind of pasta that no one else in the
(42:32):
world knows how to make. Guy in somewhere in Africa
that a certain story has been passed down for thousands
of years, and he has to figure out who's the
next person to pass it down to. It's a super
super interesting book, Let's see. I have also been on
the hunt for free public APIs for demos, and there
is a one that I like that is I can
(42:56):
has dad joke.
Speaker 5 (42:58):
Yes, thank you, I have actually used that. I'll give
you a real quick backstory. I did a video for
view Mastery a few years ago and the way that
they asked everybody to do a demo project. So what
I did was a view app that randomly pulled from
that API and then and displayed a dad joke on
the screen. So yes, that's great.
Speaker 3 (43:18):
So that's my second pick.
Speaker 1 (43:21):
Awesome, all right. If people want to We've talked about
like the discord and the YouTube channel, but if people
want to connect directly with either of you, how do
they find you.
Speaker 3 (43:30):
I'm on Blue Sky, Claire Libori and LinkedIn.
Speaker 4 (43:36):
Yep, and you can find me at Eric C.
Speaker 1 (43:38):
H E. R I. K.
Speaker 4 (43:40):
Ch on. All the socials are just Eric Hanschett on
LinkedIn as well.
Speaker 1 (43:44):
Awesome. All right, well, thanks for coming, thank you, wrap
up until next time. Max Out
Hey, folks, welcome back to another episode of JavaScript Jabber.
This week, on our panel, we have Steve Edwards Yo yo.
Speaker 2 (00:13):
Yo, coming at you aj style from just outside Washington, DC.
Speaker 1 (00:19):
Yeah. I was trying to come up with your name
and I was looking at Eric, so my brain kind
of went wait, wait, wait wait. I'm Charles max Wood
from top Endevs. And this week we have two guests.
We have Eric Hanschett, Hello, Hello, welcome back, and we
also have Claire. I didn't get your last name, Claire
well gory Oh cool, sounds Italian, so you're probably awesome, all.
Speaker 2 (00:43):
Right, So I sounded French, but that's just me.
Speaker 1 (00:48):
It could be anyway. Do you guys want to just
introduce yourselves real quickly? Got Eric on before, but Claire,
you're new to us, so we want to know how
awesome we are.
Speaker 3 (00:56):
Sure. I'm Claire. I'm a senior Principal engineer at as
in agentic AI, and I work on all seeing developers
and agents basically hero ID and the strands Open Source
Agent SDK.
Speaker 1 (01:14):
Cool.
Speaker 4 (01:16):
Well, and I'm Erica. You've been on here a few times,
a senior developer advocate at AWUS Huge VIEWJS fan it's
probably the last few times you may have heard me
on this podcast talking about it and all the work
I've done in that community and in the JavaScript community.
So glad to be here today and talk about some
really exciting that we released recently. That Claire can talk
(01:38):
a lot about Kiro, which is our new AI I D.
Speaker 2 (01:42):
So a little bit of background before we get started.
So we had Eric on a month and a half
ago somewhere around there. I forgot and found out later
that he had been planning to talk about Kiro, but
it turned out it wasn't quite ready for release at
that time, so we sort of scrambled and talked about
other things sounded we sounded really good thought. But now
Kiro's has been released, and I know I've heard a
(02:04):
lot about it on other podcasts, you know, in other
blogs and so on, and people have been really raving
about it. So we're finally here to talk about it
officially and all it can do for us.
Speaker 1 (02:16):
So I'm curious. Sometimes it's obvious why people name stuff
what they name it. Right, It's like this does email,
so we called it email thing? Right? How did you
come up with kiro?
Speaker 3 (02:28):
That was a AI assisted process? Actually? It actually means
crossroads in Japanese, and so we were looking around for
a word that would reflect the sort of new development
style that we have now. Our development has changed so much,
so we were at a crossroads.
Speaker 1 (02:52):
Gotcha. It's funny because I go for walks and I
talked to Grock the whole time, so I know how
this goes. Sometimes it's like, oh, that's a good idea,
So do you want to explain what it is and
how it works, maybe how it's different from some other
things that it sounds like do the same thing.
Speaker 3 (03:08):
Well, Kiro is an AI powered ide like Eric said,
but one of the things that makes it different is
that we've all been kind of vibe coding for a
bit now, and one of the things that we were
noticing is that people really struggled to get what they
wanted out of vibe coding, especially when it came to
(03:29):
more complex tasks. And there was a kind of an
aha moment for me when we and was the reason
that we started Kiro, which was what if developers instead
of saying kind of commands to an AI assistant of
how to change their code, what if they just declared
(03:49):
what they wanted? What if they defined requirements and then
worked with the AI on requirements and how that broke
down into individual tasks. And that was really the birth
of what we call spectrum and development, which we have
dedicated UX around in KIRO. That's really what makes it different.
Spec based development is kind of what it sounds like.
(04:11):
You create a specification for what you want. That could
be a brand new application, it could be a feature,
it could be a massive refactor in your code base.
But you go through this process where you define the requirements,
the use cases, user stories. You can have the AI
generate all of those, because I remember writing these by
(04:31):
hand ten fifteen years ago and it wasn't that much fun.
And then you go through a process of working through
the technical design with the AI and finally what tasks
that breaks down to, and then you have a lot
of control over steering the actual implementation of those individual tasks.
And we've heard from folks that this has enabled them
(04:53):
to at the end of the day, produce much better
quality code and actually get to what they wanted. I
said of this very frustrating experience with vicuding, we were
hearing about where it's like you go ten cycles and
that it's almost like the AI is not getting it
getting what you want.
Speaker 1 (05:12):
Yeah, I've experienced that a little bit. I've been using
cursor and a co pilot. I'm depending on what i'm
working on. And you talked about it breaking it down
into lists, And it's funny because cursor will give me
a these are the things I'm going to do, and
then it'll check them off as it goes right. And yeah,
it doesn't always get it right. If it's something simple,
probably the vast majority of the time it gets that right.
(05:34):
But yeah, the second it gets complicated, it's like, Okay,
this isn't working. Sometimes I'm still not quite sure what
it did or how to fix it, and so I'll
ask it to fix it, and then it still doesn't work. Yeah,
So I get that. It sounds like this is a
little bit more in depth than sort of the process of,
you know, giving it a single prompt and then having
(05:55):
it come up with I'm going to do this, and
then I'm going to do that, and then I'm going
to do that and then I'll be done.
Speaker 3 (06:01):
Yeah, and it still starts with a single prompt. But
one of the things that I've noticed is that it's
really hard for us to describe what's in our head
from the get go, from a blank prompt text input box.
And so often what I find is that there's this
iterative process of effectively the AI trying to guess what
I meant by this simple sentence and then being able
(06:25):
to iterate on that and say, no, no, that's not
exactly what I meant. I actually meant this, or oh,
that's an interesting idea I didn't even think of. And
so that process of creating this spec is really helpful
for refining what it is you want. Often if you
go to Vibe coding, it's kind of like you know,
when you have an engineer who's a bit a bit fresh,
(06:48):
and they always jump right to writing the code before
actually thinking about the problem. That's kind of the Vibe
coding experience is the AI jumped into writing the code
before kind of understand the problem. And so this spec
driven development helps you to kind of make sure you're
on the same page and that you know everything that's
(07:08):
in your head of what you're thinking about the application
or the future is actually out on paper without necessarily
having to type it out all yourself. You kind of
just go through this iterative process of the AI kind
of guesses writes the spec for you, and then you say, no, no,
that's not exactly what I meant.
Speaker 1 (07:26):
Yeah. It sounds a little bit like the process I
went through more as a freelancer than really is a
full time employee, where you know, I'd have the client
come to me and say, hey, we want this feature,
and then I'd ask questions, do you want it to
do this? Do you want it to do that? And
you want it to look like this? Do you want
it to blah blah blah, right, and then there was
further iteration once we got going where it was okay,
(07:46):
I have a rough version of this, and then they
go look at it. Oh, no, this isn't quite what
I wanted. And so by the time we're done, yeah,
I had a fully fleshed out understanding of what they
wanted and could build it. And it sounds like that's
what this is doing.
Speaker 3 (07:58):
Yeah, And I love that idea of the rough version
of it, is this what you wanted? Because so often
we don't know what we want until we actually see
it when it comes to application development. And you know,
we've seen across moving from Waterfall, where we tried to
you know, have all the requirements ironed out at the beginning,
(08:19):
and that never really worked because those were not Those
were never the actual requirements that you ended up with
at the end once you saw the application. We went
to agile, where we're attempting to move into sprints, where
we have some capability of refining those requirements as we build.
But even with something like agile development, it still takes
(08:42):
so long to get to something that's working that you
can actually show to a customer, show to a user.
So it's going to take multiple months. And one of
the powerful things about AI assisted coding is you can
get to that tangible result so much faster. You can
build entire application may not work exactly right, it may
(09:02):
not be exactly what you had pictured in your head
to begin with, but you can actually play with it.
You can you can touch it, use it, and go
back to your requirements and say, oh, actually this this
part wasn't exactly right, or oh we forgot this, let's
change these a little bit and regenerate. And so that's
been really powerful for getting into kind of prototype driven development,
(09:26):
and I've found that that ends up with a better
product at the end of the day because you are
able to really really use the product very early on
while you're still forming the requirements.
Speaker 1 (09:37):
Really yeah, I'm kind of imagining how this flows. I'm
a little curious once it has the spec if it's
something that's a little more complicated. One of the things
that I've experienced more with Cursor is that it will
go and it will do the whole job right, and
so it gets done and I'm like, I have a
(09:59):
lot of stuff.
Speaker 3 (10:00):
Yeah, yeah, yeah.
Speaker 1 (10:02):
Does it do that or does it kind of so
march through a piece at a time there too.
Speaker 3 (10:08):
We call that supervised versus autopilot mode. So sometimes you know,
you just want to get something done right. You just
want a really simple tool and it doesn't need to
be exactly right. And I find that autopilot mode is
going to be the fastest way. That means exactly what
you just said. It's just going to plow through the work.
(10:29):
It's going to completely build the entire thing. Supervised mode
is where you want to give feedback one step at
a time. So inspec mode, what you can do is
it generates a task liss similar to what you were
saying that the AI says, here's what I'm going to do,
but it will actually stop for your feedback after each
(10:52):
individual task, and so you can say, whoa, whoa, you
are starting off on the wrong foot here, let's rewind
and get this right. And I find that that does
end up with more of the code that I already
wanted because I'm steering it along the way. In between
these little individual tests. I will say that I've noticed
(11:19):
AI models really like to do a lot of work.
And so what I find is when they when they
generate these spec files, it's following all the software engineering
best practices. It's got high unit test coverage, it's got
integration tests. It might sometimes say I'm going to try
to deploy this before before deciding I'm done. And that's
(11:44):
really great if you are building a production app. Sometimes
I'm just building a little side project or I'm building
a little utility that I just kind of need to
get done and I don't need integration tests. So being
able to review that list of tasks for it goes
off and does a lot more work than I wanted
it to do is also really helpful.
Speaker 4 (12:07):
Nice I feel like it's like an eager intern. It
just wants, it wants to help you so much and
do everything for you, and you can definitely be like, no,
let's supervise this one at a time. YEA, let's list
delete things we don't need. We don't need three thousand
integration tests.
Speaker 1 (12:25):
Definitely. I love the eager intern analogy. You sent me
for coffee and I came back with doughnuts and sandwiches.
Speaker 3 (12:32):
And the whole coffee machine did.
Speaker 1 (12:35):
Your dry clean. So so do you use it now
or go ahead?
Speaker 2 (12:43):
I was going to say so in reading through the
instruction chareing what you were just talking about. It sounds
like there's a lot of it's like dock writing, almost
like you're planning and figuring out I'm going to do this.
How much of it? How much does it actually write
code versus just doing the planning? Is it all of
the above? I mean could say, for instance, if you
are a design challenge back in engineer like me or
(13:06):
a code writer like me who has no design chops whatsoever,
you know, and I actually have a project I'm working
on that I really would love to do, this can
do something like come up with at least a somewhat
workable UI design based on what you're doing, you know,
in terms of page layouts and elements, and then you
naw bar structured tables formed that kind of stuff as well.
Speaker 3 (13:28):
Yeah, it can both write the specs and write the
actual code for you. And the technical design step is
really helpful for some of the UX as well, and
kind of the code architecture, so you can say, what
front end framework do you want to use, what kind
of component library do you want to use? You know,
compared to I mean when I first started out in
(13:52):
the industry, I was literally writing waterfall requirements documents. What
I like is that it's not so much about you
writing these documents as they are kind of artifacts for
a shared understanding between you and the and the AI,
and so you don't have to actually write any of it.
(14:13):
You can keep prompting that, you can keep prompting in
KIRO until you have the requirements that you want, or
you can you know, kind of edit them yourself where
need be. But it's really helpful and really the whole
life cycle of kind of ideation and then design and
then actually writing the code.
Speaker 1 (14:33):
Yeah, I'll just jump in. I assumed that I guess
I've been working from the assumption that it's as capable
as any of the other systems out there, right, So,
like I said, cursor and copilots seem to be able
to do the UI stuff just fine, because I have
the same problem, right, I work almost exclusively on the
back end. In fact, for my full time job, I'm
basically building APIs and then somebody else is doing all
(14:55):
of the front end work. And so when I'm working
on my own projects, yeah, I tend to let it
do the UI UX stuff and it does fine. I mean,
it's not always beautiful or whatever. But the thing is
is I can more easily than tell it I want
this over here, or hey, this isn't working quite the
(15:17):
way I expect. And a lot of times again back
to the point of you have to see it before
you know what you want. Like I'll go in and
I'll just do something instinctively and I didn't tell it
I needed that, But then I'm like, oh, I need
it so that when I click here, it does this
and so. And it's usually pretty good about that.
Speaker 3 (15:36):
And you can also give it. You can give it
kind of a Napkin diagram of what you want the
UI to look like, and it will actually translate that
into requirements or design statements. So that's pretty cool as well.
Speaker 1 (15:51):
So one thing that I'm wondering having used. Like I said,
I've used Cursor and I've used Copilot. I'm working on
figuring out how to get Copilot to work with Emacs.
So I like that a little more than Visual Studio code.
But Cursor's based on DS code. Is that is that
what you did with kiro or is kiro something else?
Speaker 3 (16:09):
It's based on code OSS similar to the Yeah, so
it feels very familiar. But we were also able to,
for example, with this this speck layout. One of our
hopes is to a little bit take the focus of
the environment away from just the code. So one of
(16:30):
the things that I noticed as we were working with
code OSM is it's interesting the way that the colors
are where code is very much the focus. So in
that I use a light a light mode, and all
everything in the UI is gray except for the code,
which is in white white background. And so we were
(16:54):
we were looking at this and thinking, you know, actually,
you don't want necessarily the code to be the focus
in your ID, and I think we're not We're not
quite there yet. I think our first step is with
this kind of dedicated UX around spec based development, where
the focus is on the spec the design and the tasks.
But I think my hope at least is that the
(17:17):
ide becomes a place it is not just about writing code.
It becomes a place that is more tailored around, you know,
defining what you want, and then the code is is
kind of an artifact that comes from that.
Speaker 1 (17:34):
MM I'm trying to imagine to myself what that looks like. Also,
you're the only person I know that uses a light mode.
Speaker 3 (17:47):
I just don't like staring at black all day.
Speaker 2 (17:52):
So can we maybe go at a high level? I
don't know how detailed do you want to get. Granted,
everybody can't really see the ID since we're not screen
sharing with you can just sort of walk some steps through.
Speaker 5 (18:04):
Let's say you're starting a.
Speaker 2 (18:06):
Project completely from scratch and you want to build some
you know, maybe something more complex than a two do app.
You know, who knows what are the steps? Because now,
for instance, if I come, I'm looking at the ID
right now, and it looks like you have two different
ways possible ways to start, it gives you the option
to be the vibe respect right, So one is chatting
and one is planning first, So it looks like those
(18:28):
are two different ways you can start.
Speaker 5 (18:31):
Am I reading that correctly?
Speaker 3 (18:32):
That's exactly right? Yeah.
Speaker 2 (18:35):
Okay, So let's say I want to start with the
vibe mode. So what am I doing or how what
are the steps I'm taking.
Speaker 3 (18:43):
That's vibe coding, as I think we're all kind of
familiar with. Now you're you're prompting the AI to generate code,
it's going to you know, if you do have any example,
any existing code in there, it's going to take a
look at that and bring it into the context. But
if not, it's just going to start writing code for you.
Speaker 2 (19:02):
Right, So there's this chat panel down here, so this
is where you start saying, hey, yeah, this is my code.
I want to add this feature that's right, and make
it look like this. This is my stack, you know,
and I'm using you with you know, a level in
the back end with nursia, jas and tailwind, et cetera.
That's my preferred stack, yep. And then it's going to
(19:23):
start generating what is it going to start generating documents
there that with the output, you know.
Speaker 3 (19:29):
Yeah, so in vibe mode, it's just going to start
writing code. It's that that very familiar experience I was
seeing before, where it's just going to jump right into
writing the code and then kind of iterate from there
of trying to get the AI to update the code
into what it is you want it to build. And
so I find that that's very useful. I find that
(19:50):
the so the Cure team built uses kiro to build kiro,
and I find kind of across the team that we
tend to jump into vibe mode if we know exactly
what we want. You know, we know the code that
needs to be written. We're just kind of steering the
AI to generate that code. And it's a pretty simple,
(20:12):
simple problem. But if we don't know what we want,
we don't know exactly what code needs to be written,
will jump into spec mode or if it's just a
more complex problem. And I think that that takes a
little bit of intuition today, It kind of takes a
(20:33):
little bit of playing, and it requires knowing what do
I think the AI is going to be good at
kind of one shot right vibe mode to get what
you want. You want it to be pretty close to
one shot. You don't want to have to do a
ton of prompt iterations. But if it, you know, you
kind of get this intuition of what is simple enough
(20:54):
for me to give to vibe mode and then anything
more complex than that that needs to be broken down
into individuals or they don't exactly know what I need yet.
It is a really good starting place for a spec.
Speaker 4 (21:08):
And maybe I'm jumping ahead here, but as we're using bimode,
as Claire said, there's it reads the files, but you
can't create steering files or kind of the same thing
you might see and other code editors like rules files
that you can set up so it follows certain patterns
as you're vibe coding back and forth. And you can
even have here generate these steering files for you to
(21:32):
start off with.
Speaker 3 (21:35):
Yeah. I find that that's really the first thing to
do if you're starting in an existing project is generate
the steering files because it creates this really great summary
of the code base and what it is and what
it does and what the architecture is. That really helps
you to save on contact space, so it's not having
to read a ton of files every time.
Speaker 4 (21:59):
Yeah, outputs better code.
Speaker 1 (22:02):
I feel like, Yeah, so a couple of other questions
that I have, I'm going to change the topic a
little bit. So both Copilot and Cursor allow you to
change the model you're using. I'm assuming that that's the
case here too. I can say I want to use
Claud and then okay, Claud's not cutting it for me.
(22:23):
I want to use GPT or is there some kind of
auto selection that you do?
Speaker 3 (22:28):
Currently the model that's available in the product is clouds
on it for.
Speaker 1 (22:33):
Okay, and I can't change that.
Speaker 3 (22:37):
That's right, But we currently in previews, so things change.
Speaker 1 (22:43):
Right, Yeah, and only get better, folks. And then the
other question I had was, so both those other systems
that I've used you get for what you pay. So
my employer is paying for Copilot, and so that's why
I'm using cursorund my own stuff is because get hub
didn't give me a good way to differentiate between credits
(23:06):
that I pay for and credits that they pay for.
So I've looked at setting up a separate getub account
just for their stuff. And then anyway, but I guess
my question is, so you always have a limit on
the number of credits or premium credits you can use?
Is that the same for Kiro? I looked at the pricing,
patient looks like that's more or less the case.
Speaker 3 (23:26):
But yeah, we have I guess four different tiers free
pro pro blust power and very similar to other pricing
plans you've seen in the space, you get a certain
number requests or credits that come with that, and then
over once you go beyond that for the month, you
can enable per request or for credit pricing.
Speaker 4 (23:50):
And since we are in preview, like Claire said, things
are still in flux, things might be changing. We have
been listening really closely. You have a discord community and
there's Reddit communities, lots of different communities we're looking at
hearing the feedback has some really great people in those
threads every day listening to that. So we're constantly listening
and so there may be updates, may not be updates,
(24:12):
but stay tuned on that and check our blog on
all the pricing details. Since it's in preview right now,
it's under a wait list. So one thing is we
had lots of great feedback when we first launched it.
We had actually so many people want to try it,
we put up a wait list and so we're slowly
going through that. Weitlist, adding more people on and then
(24:32):
we'll have more information about that soon.
Speaker 1 (24:36):
So one other question that I have with this, I
feel like i'm hogging the mic, but I'm going to
keep asking my questions. So I'm kind of imagining. So
the Vibe mode sounds a lot like what I've seen
other people doing with the other agent systems, right, and
so you know, I don't necessarily see a major advantage
one way or the other between this and those. But
(24:56):
then you've got that spec mode, which is something that
is different, right, And so I'm I'm trying to think
to myself, Okay, when will that really really deeply pay off?
Is it those super complex features or tasks? Are there
other things that I'm just not even thinking of where
you're going to say, oh, well, we tried this with
(25:18):
our you know, with our workload, and it made a
major difference here and maybe not so much of a
difference on some of the other stuff we did.
Speaker 3 (25:26):
I mean, in my experience, I would say I get
higher quality code at the end of spec mode. With
five mode, I find that either I have to put
some of the practices in steering rules, like you have
to read, you have to write tests, you have to
run those tests. I find that at the end of
(25:48):
the spec mode, the AI has already done that for me.
And like I was saying, sometimes that's overkill. When it
puts that in the task list, it says, you know,
I'm going to write tests and run tests at every
step of these data. And sometimes I say, no, it
just remove all the tests. But but when I'm writing
something for production, I find that I get higher quality.
I find that I don't have to iterate on the
(26:11):
code as much once it's finished, especially if I'm steering,
just steering it along the way, And so I find
I end up with I guess the right thing the
first time more often. And I would say the other
one is I'm trying to think of We have some
examples from the Cure team using Kiro to build Hero,
(26:34):
especially in UH some of the code oss UH code base.
This is a massive ten year old code base, and
there's some changes that we've had to make in the
in the in our our fork of it. It is
(26:56):
really hard to understand that. And so I was thinking about,
you know, so many developers having to work through a
massive legacy code basis, which we know there are many
of in the world, and having to trying to make
a change to one of those code bases. This is
like incredibly scary thing, right, and our experience was that
(27:17):
that process helped us to work with the AI to
build an understanding of where those changes needed to be made.
And that was because it was able to kind of
have this planning stage where it looked through a lot
of the code, summarized, a lot of the code, figured
out all the places that needed to be changed for
(27:38):
a particular feature, and that was helpful for the developer
learning as well. Right, they didn't they didn't actually know
where where all these changes needed to be made, so
it was really helpful for that, you know, both the
steering files that were generated for that code base, but
also just going through this planning phase was really helpful
for the developer doing that.
Speaker 1 (28:02):
Cool.
Speaker 2 (28:05):
Well, yeah, I mean that would make sense, right that
anything more you put into the planning, the better your
execution is going to be on the back end, while
at the same time understanding you know, I just am
working through a large project at my day to day
and no matter how much you plan, there's stuff that's
going to change as you go along. You're going to
find out a library doesn't work or doesn't have what
(28:26):
you need, or here's a better option, or you have
to totally scrap it because of other considerations. So I
guess if kiro can handle the changes, which it seems
like it can, where you can go and say, okay,
this change, now I need to rescope this and have
this then yeah that's awesome.
Speaker 1 (28:45):
Yeah. And the fact that you can steer it as
it barfs out all that code.
Speaker 2 (28:53):
Yeah, bar is that a technical term?
Speaker 3 (28:56):
Trap?
Speaker 1 (28:58):
It is when I use cursor.
Speaker 2 (29:00):
So just from a training standpoint, I've been poking around,
you know, just on YouTube. If I search, I'll see
a bunch of stuff on hero. Do you guys have
any particular training videos or step in your channel.
Speaker 5 (29:13):
On kiro on like video stuff?
Speaker 2 (29:16):
Okay, here's how you do this.
Speaker 4 (29:19):
I could jump in here in this one.
Speaker 3 (29:21):
Yep. Yeah.
Speaker 4 (29:22):
So the full developed Rabits team, We're really excited. I've
been creating a lot of content. We have our own
Twitch channel, so we've been creating weekly every other week videos.
Claire was just actually on one couple of weeks ago,
I believe, and we've been creating. We have a YouTube
channel for kiro too. We have a bunch of videos there,
(29:44):
and we have blue Sky and x account. So yeah,
we kind of hit all the channels up to try
to get out and also just going to the kiro
dot dev website and then looking at the blogs in
the docks a great way to start.
Speaker 3 (29:58):
And the discord community is so active. That's been one
of the the nicest surprises almost of this lunch. You know,
we wanted a way to connect with individual developers, right,
but I've been so excited to see really a community
(30:19):
forming and people helping each other on the Discord. So
that's been amazing to see.
Speaker 1 (30:27):
Yeah, I have to say, with most of the things
that I've worked on, when the people are great, it
makes it just that much better. Like it, it's fun
to deliver something cool and it's fun to see that
it made a difference. But when you're able to connect,
I don't know, at least for me, it's a whole
different thing. Yeah, So where does this go next? Like,
(30:51):
what are you putting into kro next? Is there going
to be a you don't even need a developer any more? Mode?
I hear people worry about that all the time.
Speaker 3 (31:00):
I mean, some things that folks have been asking us for,
especially in the Discord, We're listening very closely, things like
background tasks folks have been asking for, where maybe you
do want to set it and forget it for a
little bit and let it go for ten to fifteen minutes.
We've been finding that that mode is useful when you
(31:23):
have a project where the AI can iterate for a
long time. So, for example, we have a project internally
that is written in RUSS that was done entirely with
AI coding, and the Rust compiler gives such good feedback
about the code that you can kind of let it.
Let the AI go for ten minutes and it will
work out all the bugs because the Rust compiler is
(31:45):
giving it such such good feedback, and so you can
let it go for ten minutes and the code that
comes out might not be you know, massive amount of
code of blab of code for you to review, but
it's gone through so much testing it's definitely going to be.
It's almost always right when it comes back. But you
(32:10):
also don't necessarily want to be you know, running your
local laptops to you for one hundred percent. For that
whole time, the folks have been asking us, you know,
how can we run it? Have the task maybe run
on a different machine, like on an easy to instance
something like that, or just you know, multiple threads going
on in their local laptop. Let's see some other things
(32:32):
that folks have been asking us about.
Speaker 1 (32:34):
I'm imagining somebody using that Kiro build me Twitter.
Speaker 3 (32:38):
Oh yeah, yeah, somebody already me a YouTube clone that
they built with Kiro. But let's see what else. There's
a lot of opportunity for improvement in the spectrum and development.
I think we're just getting started there. We have some
really interesting ideas about how to prove that the code
(32:58):
that was generated actually meets the requirements that you set.
I think we've all seen that. Sometimes models will say
that they have met your requirements and they have not
met your requirements, and so we have some really interesting
ideas there.
Speaker 1 (33:13):
Yeah. I've experienced that too, where it says I did
it and I tried it and it worked and no.
Speaker 2 (33:21):
All right, so Eric, what is the Sorry? Go ahead, we're.
Speaker 1 (33:25):
Getting close to the time for picks, so when you
to start writing down, But that's fine, go ahead and
ask your question.
Speaker 2 (33:31):
So, Eric, what is the AWS channel on YouTube or
what is the channel for Kiro?
Speaker 4 (33:36):
Yeah, there's a Kiro channel on YouTube. I think it's
kiro dot dev. So if you go to YouTube dot
com slash at kiro dot dev you.
Speaker 3 (33:45):
Will o t k Dev.
Speaker 1 (33:47):
Yeah, thank you.
Speaker 4 (33:48):
Yeah, k I R O d O T DEV. That's
a good call it yep.
Speaker 1 (33:53):
Okay, And do we get the visuals on there? What's that?
Do we get to see your faces on there or
somebody else if.
Speaker 4 (34:03):
You look there today? As of this recording, I'm not
sure when this is going out, but I said today, Yeah,
I just did a video with James Ward. It's extending
tirost capabilities with MCP. We didn't even get into this,
but it does have MCP capabilities, which is I'm sure
you guys done shows on it. But yeah, we go
through in this in the latest video in the channel
on how to I add some fun MCP servers how
(34:26):
to set that up? So yeah, and there's a lot
of interviews, a lot of great information on there to
check out.
Speaker 2 (34:31):
Wow, when you're dressed up too, it looks good.
Speaker 4 (34:34):
Yeah, I got I got to wear the sports tacking occasionally,
not always just wearing the.
Speaker 1 (34:39):
Black on black.
Speaker 2 (34:40):
Look, that's really good man.
Speaker 1 (34:44):
All right, Well let's go ahead and do some picks, because,
like I said, I have a hard stop. So we'll
go ahead and start with Steve. Steve, what are your picks?
All right?
Speaker 2 (34:54):
So don't have anything else outside of the dad jokes
of the week, So we'll dive right into those. So
someone threw a can of sodat maybe the other day,
but I'm okay, it was a soft drink. Simple question.
A couple of questions here. Why do hummingbirds hum because
they don't know the words.
Speaker 3 (35:16):
Thank you?
Speaker 2 (35:18):
And finally another question, why do people say tuna fish sandwich?
Nobody says chicken bird sandwich. Those are the dad jokes.
Speaker 3 (35:31):
Of the week.
Speaker 1 (35:35):
All right, you said you didn't have anything else, that's correct.
All right, I'm gonna go next. So Claire, you haven't
been on the show before, so I'll just explain. We
just shout out about whatever, and so Steve shares the
best of the dad jokes, and then you know, we'll
do other things. So I usually pick up board game
(35:56):
or a card game, and so I'm going to pick
a called The Infiltrators. Infiltrators came out in twenty thirteen
or twenty twenty three, sorry, and it's done by Pandasaurus.
They have other games. It is a two to five
player card game. And what it is is you have
a certain number of I guess spies that you're trying
(36:19):
to or mafia members that you're trying to pinpoint. And
so what they are is they're face down cards in
front of each player, and so as you look at
the there's a delay in my videos striting crazy. So
as you play, then what you can do is you
can so I can give a clue on my cards
(36:42):
or somebody else's cards, but if I play it on
somebody else's card, then I get to draw a card
back to my hand. Otherwise I have to take a
turn to get more cards eventually. And so usually you're
cluing other people's cards, but sometimes you clue your own
because you know what yours is and nobody else does,
so you know. Then you can kind of give people
the you know, instead of them guessing over and over again,
(37:04):
you can just basically tell them what it is based
on what they already know. And so if I put
put a card on somebody else's they can turn it
one way or the other that says there's nothing in
common with this card, or there's something in common with
this card, and it basically what it can have in
common is either the numbers on the card. And so
let's say that you have the fifteen that you play
(37:26):
over there. Well, the fifteen shows all of the factors
or multiples of the number, right, So if you play
a three as three, six, nine, twelve, and fifteen, if
you play the fifteen, it has one, three, five, and fifteen.
And so if my card is a one, three, five
or fifteen, I don't think there were ones anyway. If
it has any of those numbers, then I'll put it
(37:47):
down and say there's something common, or if it's the
same color as something common. But if there's nothing in common,
then you can eliminate a whole bunch of cards. You
can you say it's not blue anyway. So it's it's fun,
and what you're trying to do is you're trying to
expose what all of the cards are. And so one
of the things you can do on your turn is
instead of putting down a clue, you can say you're
(38:11):
supposed it has a little gun prop and you're supposed
to execute the card. But we always just say that
one's the three of red, and then you know, whoever
had it flips it over and says you're right, or
if you're wrong, then there are consequences for that. But anyway,
it's a fun game. Board game geek waits at at
two point one point one. That's out of five, and
(38:32):
so it's it's pretty approachable for people for the casual gamer.
Sometimes I like the deep dark games that are like
three or three point five and weight. But you know,
if you just want to pick up a game you
can play in a half hour, that's a great one.
And it's just it's a deck of cards, so anyway,
not a like a a poker deck, but it has
(38:56):
its own deck. But anyway, super fun. And so I'm
going to pick that we played it, I don't know,
like six times the last time I got together with
my buddies. We get together and we play every Wednesday.
So we're going to play tonight. We'll see what we play.
Speaker 4 (39:11):
What's the name of the game again.
Speaker 1 (39:13):
Infiltraders, Traders is spelled like traders like people who betray people,
so it's not the typical spelling of the word. And
then I'm also going to pick While I was in
Atlanta last week, I was in the hotel trying to
just relax, and so I was looking for a show
to watch and I picked up The Twisted Tale of
(39:35):
Amanda Knox and Amanda Knox is an American student who
was studying in Peruja. When I was a missionary in Italy,
I didn't live that far from Peruja actually anyway, and
that's been kind of fun for me because like half
the show's in Italian with subtitles because the Italians are
speaking Italian. But anyway, so she gets arrested for murder
(39:56):
and so they're kind of going through the trialing their
five episod So it's out as we record this and
it comes out every Tuesday or Wednesday. But yeah, really
really enjoying that, and so I'm going to pick that show.
I've also been watching The Terminal List and that has
Chris Pratt in it. That one's a little bit more violent,
but I'm enjoying that one as well. Twisted Tale of
(40:16):
Amanda Knox is on Hulu, which is included in our
Disney I've been watching on Disney Plus, and then a
Terminal List is on Amazon Prime. So yeah, I'll pick
all of those. I'm trying to think what else. But
I'm going to be speaking at commit Your Code conference
in at the end of September, and I believe this
will come out before then. So if you're going to
(40:36):
be there come find me. And then on the next
week I'm going to be at Rocky Mountain Ruby Conference
in Boulder or no, that's two weeks later, but anyway,
so I'm going to be around if you're in Atlanta.
Price Picks is based in Atlanta, so I wind up
out there. So just email me and say, hey, I'm
in Atlanta. Let me know when you're coming out next
and I'll let you know. But yeah, those are my picks, Eric,
what are your picks?
Speaker 2 (40:58):
I good?
Speaker 4 (40:59):
Well, So last night I went to the movie theater
with my wife and I guess Hamilton has a ten
year anniversary movie that you can watch, So we watched
the ten year Anniversary Hamilton movie. That included like ten
fifteen minutes the beginning of like interviews with the cast
and everything. So that was kind of a fun evening
(41:20):
to check it out. Of course, you can get Hamilton
and Disney plus the movie, but it was nice to
check that out in the movie theater, which I've never
watched Hamilton. I've watched it at in Broadway, but I
have not watched it in a movie theater, which so
that was fun. So that's my first pick, and then
I guess the second pick would be I just went
to Singapore for an AWS event a few weeks ago. Man,
(41:42):
I love Singapore. I just love the food, the hawker stations,
the all the amazing gardens. So I highly recommend people
who want to travel to a nice destination. Have I
went to Gardens by the Bay, which had this crazy
light show. They do it like once every day or
twice a day, and so yeah, great weather, great food,
(42:05):
highly worth checking it.
Speaker 1 (42:08):
Awesome, Claire, what are your picks?
Speaker 3 (42:11):
Let's see. So I just finished a book called Custodians
of Wonder by Elliott Stein. Ancient customs, profound traditions, and
the last people keeping them alive. So this guy went
and found people who are kind of the end of
the line. There's a woman in Italy that makes a
certain kind of pasta that no one else in the
(42:32):
world knows how to make. Guy in somewhere in Africa
that a certain story has been passed down for thousands
of years, and he has to figure out who's the
next person to pass it down to. It's a super
super interesting book, Let's see. I have also been on
the hunt for free public APIs for demos, and there
is a one that I like that is I can
(42:56):
has dad joke.
Speaker 5 (42:58):
Yes, thank you, I have actually used that. I'll give
you a real quick backstory. I did a video for
view Mastery a few years ago and the way that
they asked everybody to do a demo project. So what
I did was a view app that randomly pulled from
that API and then and displayed a dad joke on
the screen. So yes, that's great.
Speaker 3 (43:18):
So that's my second pick.
Speaker 1 (43:21):
Awesome, all right. If people want to We've talked about
like the discord and the YouTube channel, but if people
want to connect directly with either of you, how do
they find you.
Speaker 3 (43:30):
I'm on Blue Sky, Claire Libori and LinkedIn.
Speaker 4 (43:36):
Yep, and you can find me at Eric C.
Speaker 1 (43:38):
H E. R I. K.
Speaker 4 (43:40):
Ch on. All the socials are just Eric Hanschett on
LinkedIn as well.
Speaker 1 (43:44):
Awesome. All right, well, thanks for coming, thank you, wrap
up until next time. Max Out
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
NFL Daily with Gregg Rosenthal
Gregg Rosenthal and a rotating crew of elite NFL Media co-hosts, including Patrick Claybon, Colleen Wolfe, Steve Wyche, Nick Shook and Jourdan Rodrigue of The Athletic get you caught up daily on all the NFL news and analysis you need to be smarter and funnier than your friends.