YT YouTube2Text

YouTube2Text → Transcripts → AWS re:Invent 2025 - Spec-driven development with Kiro (DEV314)

AWS re:Invent 2025 - Spec-driven development with Kiro (DEV314) — Transcript

by AWS Events · 8,813 words · 1,288 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Hello everyone. Thanks for making it.
0:02Today we are going to be talking about
0:07spec driven development with Curo.
0:10This is going to be a
0:14we're going to be doing a code talk here
0:16today. So I have my laptop set up. We're
0:18going to look at some code, but we're
0:19going to first go through some slides
0:21and I'll explain what we're up to.
0:23First, let's have a a little bit of an
0:25introduction. My name is Eric Hchett.
0:27I'm a senior developer advocate at AWS
0:30and today I have with me Nick Hill.
0:33>> Hey folks, I'm Nick Hill. Uh I lead the
0:35product team for Kira. Uh thank you all
0:38for coming here. This is after lunch.
0:39We're going to have a fun session. Hope
0:41you you won't need your coffee. Make it
0:44fun.
0:46Absolutely.
0:49So we're going to start with a few
0:51things. We're going to look at the
0:53history of Curo. So Nuel is going to
0:56jump directly into to that in a moment.
0:59We're going to talk about a little bit
1:01of spectriven development, the name of
1:03this talk and why I think everyone here
1:05should be using it with their agentic
1:07ideides today and we're going to
1:09specifically look at Curo and how that
1:11works with it. We're going to touch on
1:13MCP. Does anybody here use MCP servers
1:16right now?
1:18Okay, I see a handful. Let me ask you,
1:20who's using Curo right now?
1:23Okay, only a few. So hopefully by the
1:25end to end of this hour talk that I have
1:28convinced you all to try out Kirao and
1:30Spectdriven development
1:31>> and and uh just another followup who's
1:34uh who's used Kirao so far.
1:37>> Okay, a few more people.
1:39>> Okay,
1:39>> good to know.
1:40>> Good to know. And then this is a code
1:43talk so I see everyone has their laptops
1:46out or most people. We are going to jump
1:48on I'm going to jump on this laptop
1:50here. We're going to do a little bit of
1:50pair programming. So, we're going to try
1:52uh to create an app. Uh we're going to
1:55talk about it more in a moment, a job
1:56seekers app. We're going to use the
1:58spectriven development flow inside Curo
2:00and we're going to see how well it does.
2:03Uh and then we're going to talk about
2:04some of the features of how it works all
2:07together. And then I want to give every
2:09everyone here time to ask us questions
2:11at the end. So maybe uh 10 15 minutes to
2:15to for that.
2:18So those of So from this the hands that
2:21came up earlier sounds like a lot of you
2:23have not used much of Curo. So let me
2:26let me explain it before we get too much
2:27further. It's a new agentic IDE that we
2:32released in the last few months. It
2:34actually just went G or general
2:36available a few weeks ago and it's we're
2:41really proud of it. It's it's really
2:43based on this idea of spectrum
2:45development. Now you can do vibe coding
2:47but we really feel like spectriven
2:49development is what it makes it sets it
2:52apart from some of its competitors that
2:54maybe some of you have heard of. Uh when
2:56we released it, it was a total surprise
2:59to us. It really went viral. It got
3:01hundreds of thousands of downloads. We
3:03had to put up a weight list. Uh and then
3:05people found a way around the weight
3:07list. So we had to put even bigger
3:09barriers up. But we've been really happy
3:11with reception. We've gotten so much
3:13feedback on it and we've been making it
3:15even more uh better
3:18uh as as it's been going on as this time
3:20frame um in the last few months.
3:25So I'm going to kick it off to Nel to
3:28talk about the history of Kira.
3:32Cool. So I'm going to spend a little bit
3:33of time talking about this, but I want
3:34to jump I want to jump to the coding
3:36session fast. So uh the evolution of
3:38building software uh let's actually take
3:41a step back and talk about the overall
3:43editor space right ids got really
3:45popular because they provided a single
3:46tool chain where you could do things
3:49like have an integrated compiler view
3:51lint linting issues find uh de debug set
3:55breakpoints and it it offered a lot of
3:58benefits over you know traditional text
4:00text edit like code interfaces and so
4:03that that was kind of the first wave
4:05that got wide adoption
4:08And then in the last I'd say two years,
4:10we've had really like AI editors become
4:12a big thing. It first started with code
4:15completions. So like you'd be in a file,
4:17you type like a hash or you type two
4:20slashes and you you'd be like do this
4:22and then it would generate like a
4:23function for you. And then that quickly
4:25evolved to the more agentic or like you
4:28know I call it like the chat GPT like
4:29experience where you you input natural
4:32language and the agent's able to edit
4:34multiple files.
4:38The fundamental shift that's happened
4:40and we see this with all our customers
4:42who are who've been using Kirro and and
4:45Q developer in the past was that uh
4:48unlike previously where the developer
4:50drives and provides I think with AI
4:53editors the developer steers the AI
4:55agent to author and review code the
4:57developer is still very much in the in
4:59the in the driver's seat but it's a
5:02different workflow where you can
5:03actually use AI to kind of accelerate
5:05your workflow.
5:07But you're still responsible for the
5:09code. That's the important bit.
5:14Vibe coding. This term got really
5:15popular in the last year year and a half
5:17where it's essentially covers that this
5:19workflow where it goes developer writes
5:21a prompt, AI generates code, you write
5:23another prompt and then you keep going
5:25in this loop and cycle. And this has
5:27worked really well for a lot of people
5:30to be able to prototype things that
5:31would otherwise take weeks to set up.
5:34But it's also got its uh challenges. Um
5:38Slack threads for example, I use Slack
5:40threads as a good analogy. So if you
5:41have a a Slack convers, we've all been
5:43in Slack conversations where like the
5:45message thread goes to like a 100
5:46messages and you know you leave that
5:49thread with like okay, what did we take
5:50away? Uh why why are we making the
5:53decisions we're making? And and
5:54sometimes that's that's just lost in
5:56that context. And that's that's one of
5:58the challenges you have with vibe coding
6:02is that it kind of completely skips the
6:04traditional SDLC life cycle, right?
6:07Historically, when we build features,
6:09we've always started with artifacts,
6:10right? Require we talk we talk about
6:12requirements, we we write design docs,
6:14we talk about trade-offs, and we meet
6:16and then and then we start. Whether it's
6:18waterfall or agile, you're you're doing
6:20you're doing that both at different
6:21scales.
6:25So what we did is we looked at we really
6:27talked to a lot of our power users
6:29internally in the company and externally
6:31and we saw that they had already made
6:33this organic shift to planning before
6:35they actually went about creating or
6:38generating code and planning up front
6:41really helped the models focus and
6:42produce better output. So what we did is
6:45we built a a firstparty workflow in for
6:49planning called specdriven development.
6:51And normally this this time I demo it,
6:54but we're going to spend so much time
6:55talking about it that I'll just move
6:56move ahead and get get to that quickly.
6:59But really specri development what it
7:00introduces it introduce a workflow for
7:03going from your initial requirements to
7:05to an output. It's a really structured
7:07workflow where you start with a you
7:10start with requirements.
7:11Once you your requirements are done,
7:13there's a design. So you actually look
7:15at your design file and then you get to
7:17a task list and then you implement that
7:18task list. And you can go back and forth
7:20in this loop um till you and end up with
7:24an with an output you're happy with.
7:29And once your code is generated a big
7:30another big problem in this space the
7:32term AI slop has become quite common now
7:36is where the AI generates a lot of code
7:38and there's you once you generate code
7:40there like there are no processes to
7:42actually
7:43a lot of tools don't actually have
7:45processes to review things. So we have
7:46hooks built in and what hooks allow you
7:48to do is we have a whole bunch of life
7:50cycle events in the IDE such as file
7:52save, file create uh and we're adding
7:55adding a bunch more uh in the next few
7:58weeks that actually allow you to
7:59validate changes. So an example of a
8:02hook that we use in our docs repo uh is
8:06is a c effectively like a c spell
8:08checker but it also checks for tone and
8:10style. Right? So our workflow when we're
8:14writing docs for new features is that we
8:16take our initial requirements doc. We
8:18feed that into curo and it it structures
8:21our docs but then that we have hooks
8:23that help actually say hey did like it's
8:25almost like a separate thread to think
8:27about the work that was created that
8:29help help us set up rules. Another good
8:33example is for if you're building UI
8:34components, you can have things like,
8:36you know, you have a React component,
8:38you want it to follow a single single
8:40responsibility principle, right? So it's
8:42a single focused, you can use a hook for
8:44that.
8:47And rather than me keep talking about
8:49this, why don't we actually just go
8:51ahead and build something together? And
8:54we're going to pair program. Though
8:55there's only one chair, so I'm going to
8:56peer over Eric while he's programming.
9:01>> Sounds good. All right, let's get this
9:03set up.
9:07Log in.
9:10One second.
9:20Okay.
9:22So, uh we want to create an app with you
9:25guys today. So before we started, me and
9:29Nikil came up with this Figma. So do you
9:32want to explain a little bit about what
9:33it is?
9:33>> Yeah. So just FYI, after this demo,
9:36we're both quitting our jobs and this is
9:37going to be our new business cuz we
9:39really like the idea. Um but the core of
9:42it is a job seekers app. So uh problem,
9:45here's the elevator pitch. So problem
9:47people looking for jobs outside of just
9:49prepping for interviews, there's a lot
9:51more than just goes into like do you
9:53know the answer. It's like what's your
9:54what's your style? What's your tone and
9:57how do you actually prep? People use use
9:59software tools out there, but like you
10:01can actually talk to another person and
10:02there's no experience where you get
10:04feedback on how how you're doing on your
10:06interviews. So that's the core of the
10:08idea and there's you can imagine a lot a
10:11we can build a lot of things around it
10:12around it. So the way Eric
10:15[clears throat] and I started this is we
10:16just hopped on a quick call and we
10:18whiteboarded a flow where we said okay a
10:21user logs in uh they set up a profile.
10:24Uh and once they set up a profile they
10:26can actually start interview prep and a
10:28very simple MVP flow would be okay you
10:31start start my interview prep and you
10:33you go in this question answer flow. So
10:37you are you're get you're asked a
10:38question, you can record your response
10:41and that goes back and forth and at the
10:44end we deliver uh an overview a
10:47qualitative overview of how you did with
10:50quantitative uh with a quantitative
10:52score.
10:54Excellent.
10:56So we have this picture that this Figma
10:59that we created and here is KO. Just for
11:02those of you who have never used it, let
11:04me give you a quick tour. I have a a
11:06brand new VIT app that I just started
11:09this morning that we can start off with.
11:12It's basically hello world. So, we're
11:14just going to keep it to the front end
11:15for now, but you can imagine some of you
11:17DevOps and and backend engineers, you
11:18can imagine we might have a backend
11:20component to this, too. Uh, I just
11:23wanted to start with this also has
11:24Tailwind CSS installed for our design
11:28graphics. Now, on the on this side here,
11:30we have this little ghost. I haven't
11:32really talked about this. This is our
11:34Curo ghost. It's our mascot. If you have
11:36been around the Venetian, we have the
11:39Kira House. Anybody seen the Kira House?
11:41That looks like a haunted house. Yeah, I
11:43would recommend everyone to to go
11:45through it. It's actually pretty fun.
11:46And then at the end of it, there's
11:48actually a code to get a,000 free
11:52credits. 500 thou 500 or a thousand. Um,
11:55so check that out. You can go through
11:57through it at any time. Uh so this this
12:00this ghost here brings up a menu on our
12:04left hand side. Agent hooks, agent
12:06steering, specs, and MCP. So hooks, we
12:09kind of explained it already. It's a way
12:11you can essentially have things work in
12:15the background that are triggered on
12:17certain events. So imagine that you save
12:19a file, delete a file, you can have
12:21things happen. One common scenario is
12:23you might update documentation on every
12:25file save or you might uh do some
12:28localization and we actually even have
12:31pre-built ones you can kind of take all
12:33take over uh without having to write
12:36anything yourself and you can also have
12:38it specify certain file types. We're
12:40adding as NIL said more flexibility with
12:44this agent hooks in the future but I
12:45just want to give it a shout out. We've
12:46also seen very creative people from the
12:48community do this in all sorts of ways
12:51for their git workflows
12:54for testing all sorts of things. Uh
12:57agent steering is more like a rules file
13:01that you can put in maybe you have
13:03certain coding standards uh or design
13:06standards that you want to include. You
13:08can put it in steering files. I can also
13:10just have it automatically generate the
13:12steering files for me. And so what this
13:15does is it looks inside this folder
13:18inside my application and then it
13:20creates a few files. You can see here
13:22it's already created the product and the
13:24tech file. And these are just based on
13:27what's in the app right now. So if you
13:28have an existing application and you
13:31want to kind of play with Curo today,
13:34you can put this in there and have it
13:36generate the steering files and then
13:37that'll get added to the context every
13:39time you start talking to the IDE.
13:43>> Maybe you can you can show what that
13:45looks like, what those files look like.
13:46>> Yes, that's a good idea. So here is what
13:48the product is. It's not that much
13:50because we really haven't added anything
13:52yet, but it's a Vue3 web application
13:54with TypeScript and Vit. I use a script
13:57setup. I don't know if anybody's a
13:59Vue.js fan, but that's essentially how
14:02they are created. The components type
14:04safe. We have the structure. Right now,
14:06it's just kind of the hello world out of
14:08the box structure, but it's nice it's in
14:10here. Uh, and then here's some tech,
14:14which I just scroll down, but yeah, the
14:17frameworks, TypeScript, V, everything
14:19you would expect in there. And you can
14:22keep this right now. This is just a
14:24moment in time snapshot of your current
14:27application. However, you can be
14:30creative with those hooks and maybe you
14:31have it continuously update those
14:33documents and those can be shared with
14:35your product teams or whoever else. And
14:37of course, uh I heard I saw some hands
14:40earlier. MCP my I'm I'm a huge MCP fan.
14:43It's getting there. It's even getting
14:45even better. I've seen some some stuff
14:46coming out recently. But this is where
14:48you would set uh your MCP servers. In
14:51this case, I have contact 7. I have AWS
14:55loves MCP servers. If you don't know, we
14:58keep continuously adding more for
15:00different services and different
15:01products. Kind of the catchall one is
15:04the docs one. I'd recommend to try that
15:07one out. Uh but if you have a specific
15:09use case, you can install like the MCP
15:11server for Bedrock agent core is one of
15:13them. And it's just a few lines. We in
15:15fact just recently uh added a way to add
15:19MCPU servers with one click. So, if you
15:22go to our curro.dev, I'll show that in a
15:24little bit. There's a little there might
15:26be a little button there. You click on
15:27it, it'll just automatically install it
15:30inside your inside kuro. And then we are
15:33diligently going to many of your
15:35favorite open- source MCP GitHub
15:38repositories and updating it to add the
15:40little oneclick install.
15:42That is still working. I'm still working
15:44on that. And then, of course, I have
15:46Brave Search, Bedrock, Agent Core. So
15:49that that's kind of a little bit of of
15:52of how of some of the features of curo.
15:55But I know the name of this talk is spec
15:57driven development. So let's let's dive
15:59into that. So once again I on the right
16:02hand side I have two boxes here. I have
16:04vibe which I I just love to vibe. It
16:06does really good job at doing that or
16:08spec uh agents. Right now we support
16:11claude sonet 45 4 um 4.5 opus haik coup.
16:17So, we kind of recommend just to start
16:20off with auto because we pick behind the
16:22scenes the best model that you should be
16:24using. I'll just leave it at auto. We're
16:27adding more and more all the time. It's
16:29definitely in the back of our heads of
16:31adding more models for those of you who
16:33are fans of other ones. Uh, sit tight.
16:36So, but let's write spec. Uh we actually
16:39had a before I well let's let's go ahead
16:43and and start it and then I'll I'll give
16:45you a little story I had the other day.
16:47So uh here is the diagram that me and
16:51NIL came up with. I'm just going to make
16:53a snippet of it a screenshot
16:57and I'm just going to paste it here and
16:59let's give it a prompt. So I'm going to
17:01choose spec here and I'm gonna say
17:03please use this picture and create a job
17:07seekers app.
17:10Use the picture as a
17:13um uh as the design.
17:18I don't know. I always I always like to
17:20be uh explicit and
17:23I like to be polite to AI too. So you'll
17:25see me using please and thank yous. I
17:28would do that too.
17:30Okay, so now it's going in and it's
17:33taking a look at those steering docs
17:34that we just created and it's going to
17:37create this job interview preparation
17:39flow. And it's starting off with the
17:41requirements document. And I'm going to
17:43have Nicole really walk me through on
17:45what his idea is and how properly you
17:49could [clears throat] look through a a
17:50requirement stock.
17:52>> Cool. Awesome. Let me let me actually
17:54switch to the preview view of that
17:55requirement stock.
17:56>> Okay, sure.
17:56>> You don't mind?
17:58>> Here, I'll do it. Okay, go ahead.
18:01>> All right.
18:05So, as you can see, uh, we took Kira
18:07took
18:08>> We're going to switch. We're going to do
18:09a little
18:10>> Yeah, we're going to play a little
18:11musical chairs.
18:12>> Mhm.
18:13>> So, as you can see here, Kira took the
18:14initial
18:16screenshot that we had and actually
18:18unpacked that into a format. And I just
18:22realized Eric has an o opposite scroll
18:24from me. So, it's like really this
18:26throws me off when I'm trying to scroll.
18:29uh but it it unpacked our initial prompt
18:31to a set of requirements and the way we
18:33define requirements is we we define them
18:36as a user story and every user story has
18:39acceptance criteria.
18:41Now this is actually something we built
18:42we were doing inhouse in in Amazon in
18:45the in our teams the way we worked. So
18:46we took the same workflow that we had uh
18:50on our team and we made it available
18:53broadly and the benefit of a user. So
18:56user story is like the human readable
18:58way that if we want to say hey do we
19:00want what's what feature do you want to
19:01support? So this one's logging in so
19:03that I can access my application profile
19:06management whereas the acceptance
19:08criteria are the actual details. So if
19:10you you were to sit and say, "Okay,
19:12let's go and build this profile
19:14management feature, what are the
19:16nitty-gritty details that I need to
19:18actually think about?" So the the first
19:21pass I typically do is when I'm when I'm
19:23building something with hero is that I
19:25will look at the user stories and see if
19:27they're the user stories that I actually
19:28want to build. Um, and I'll start there
19:31before I get into the acceptance
19:32criteria. So if we start at first one's
19:34about logging in,
19:37the second one's about profile
19:38management.
19:41The third one's about an interview prep
19:43session.
19:47This is about viewing and responding to
19:49answers. And like this is very
19:52important.
19:55Okay. This is about navigating through
19:57multiple questions.
20:01This is about the analysis and the
20:02result.
20:06This is about trend analysis.
20:08Requirement seven.
20:10Requirement eight is about requirement
20:13eight is my least favorite requirement
20:14because it's really nothing. Like what
20:16do you mean I have a clean and intuitive
20:17interface? Uh so what I'm going to do is
20:21actually like simplify this cuz I've I
20:23just thinking through this I think that
20:24demoing all here is not that important.
20:27So, I'm just going to say um please
20:30remove requirements
20:33one, two,
20:38uh we want three, we want four,
20:43we want five, we want six. I don't think
20:46we necessarily care about state
20:48persistence either right now. We just
20:50want to get something up and running.
20:52seven and eight
20:55as they are not needed for our MVP
20:59scope. Look, I use please as well. I
21:02never use please.
21:04I'm telling you. Although some people do
21:06say it's just waste tokens, but I don't
21:08know
21:11the and this is something I really think
21:13even if you're not going to use Curo
21:15after this talk, just being able to have
21:18your ID or having your IDE AI IDE like
21:22go in uh and
21:25create these documents, but then also
21:27kind of going back and forth because
21:29oneshotting this may not give you
21:31exactly what you expect at the end of
21:33the day. So it's it's almost like a you
21:36kind of go back and forth process.
21:38>> I think the way I've seen seen actually
21:41talking to some customers is that you
21:43know the initial whiteboard we did right
21:46>> some if like as a developer I might
21:48prefer to just uh vibe code that right
21:51like just cuz then it forces me to think
21:53about my interfaces and things like that
21:54that I'll need and then enter something
21:57more structured. So that is also a
21:58workflow that uh we we we support right.
22:02So I think hopefully these requirements
22:05look good. I think for the interview
22:07questions we should just limit for
22:09interview questions limit to three.
22:16Yeah. So it should be just a quick I
22:18imagine it you go in it asks you a
22:21question and then you use your
22:22microphone explain it then it goes to
22:25the next question and then it does an
22:26does an analysis at the end.
22:29>> All right. I I think this is good. Um
22:36yeah uh so now that the requirements are
22:38done we can go ahead and move to the
22:41design phase. So what what's going to
22:42happen now is that it's going to take
22:44these requirements and generate a design
22:45dock based off your existing code base.
22:49And Eric and I will play mult uh musical
22:52chairs again.
22:54And for those of you who are thinking
22:56well my PM does this right I don't have
22:58to write requirements. Trust me, maybe
23:00you take what they give you to put in
23:02here so you can then really create the
23:05feature or product that they're looking
23:07for.
23:10So yeah, now it's creating the design
23:12documents, kind of reading what we've
23:13done and the updates, the requirements
23:16to go to this next phase, the design.
23:19Uh, while it's working, I can also click
23:21here to kind of get an idea of of the
23:24state of it. Uh, while it's working, one
23:27thing I was going to say with the spec
23:29and the vibe code, I played a game on
23:32stream the other day. I get to I get to
23:34have so much fun. We we actually had
23:37another person on and we decided like
23:39which task would I use vibe coding
23:41versus spec mode. And it's actually
23:43harder than you would imagine to figure
23:45out which one is is which one you should
23:48use. Uh, I usually say if it's something
23:50that needs a little bit more
23:51forethought, it's a larger feature, then
23:55spec mode is good. And I also I love
23:57just having this documentation
23:59afterwards I can look at.
24:01For those of you pink, for those of you
24:03eagle-eyed on the screen right there, it
24:05says, can, by the way, can everybody see
24:07this screen? Do I need to make the font
24:09bigger in the back? Can you see it?
24:12Okay, I see thumbs up. Okay.
24:15>> So what it did now is actually it
24:17created something called property tests.
24:19I want to
24:20>> Yeah. Why why don't we start with the
24:22just just the architecture, right? Like
24:24it just
24:25>> work down to see if it's the actual
24:27system we want and then I think let's
24:29get to properties
24:31and
24:31>> yes that's a good point. I'm so excited.
24:34I want to jump into property testing but
24:36we'll wait on that. So here is the
24:38design document that it created. It also
24:41gave me a nice summary of what it did uh
24:44and and I can move on from there. And
24:46it's asking me, hey, does this look
24:47good? We can move to the next phase. But
24:49I can look here and kind of take a look
24:51at the highle architecture. We have view
24:55uh nice little asky art here. But we do
24:57use mermaid diagrams. If those of you
24:59are familiar with that, it works really
25:00great in markdown. So we have start
25:02screen question results. Okay, I'm
25:04assuming this looks good. I mean just it
25:06does look good. Local storage to to
25:08share everything. web audio API that'll
25:11really help us when we record the audio
25:12from us when we're answering these
25:15questions.
25:17Uh this is kind of how I at Amazon we
25:19are a AWS we're a very document-driven
25:23culture. I've been in many design doc
25:25reviews and we actually write our design
25:27documents very similar to this. We have
25:29interfaces that we outline. Uh if you
25:32ever want to see how kind of how the
25:36sausage is made when you um want to see
25:39how some of our services are designed,
25:40kind of take a look at these design
25:42documents. So I have question response
25:45the component start screen view.
25:47>> Uh so just under response, can you just
25:49scroll up for a second? What is session
25:52state capture right now?
25:54>> Questions responses.
25:56>> Okay, gotcha.
25:58Yes.
26:00Oh, and then we have question
26:03screen.view.
26:04So, think these are all our components
26:06that it's going to create. Our score,
26:09some prompts, question bank.
26:12It's going to be just a hard-coded file
26:14for now.
26:15>> So, is this is this the actual file or
26:17is it just going to use generate its own
26:19questions?
26:20>> It looks like uh it's a static array in
26:23the application for the MVP. We'll use a
26:25predefined set of common interview
26:26questions. So I'm assuming this is
26:28>> just the structure.
26:29>> Yeah, the structure of it, but could be
26:32very similar to these or these. Here's
26:35our stored session. And then the last
26:38part we'll come back to it, but that's
26:39the correctness properties.
26:42One thing I I noticed out of here that's
26:44missing is it doesn't really have any
26:47information about how if we're recording
26:50audio from our laptops, how is it going
26:53to
26:55take that audio and transcribe it? So I
26:57want to add that feature into it.
27:00>> Yeah. And I think there's one more thing
27:01missing too. So once you transcribe it,
27:04how is it actually like how are you
27:05going to get the results? Was that was
27:06that captured?
27:08>> Yeah, that's a good point. So results,
27:11we have a session. We have a response.
27:13>> Guessing we're going to use AI for the
27:14result generation.
27:15>> Yeah, we have a results screen with a
27:17score number, but it's not really like
27:19this isn't really showing exactly what
27:22we want, right? We want a little bit
27:24more information.
27:26Sometimes I just do a quick search, too,
27:28if I can't find it. Yeah, test complete
27:30results. All right. So, what I assume we
27:32probably want to add in here is a way
27:34for AI to go in and and do some
27:38analysis. So, let's add these two
27:40things. We'll say, "Please transcribe
27:44the text as the user speaks it." Also,
27:50let's
27:52add additional analysis.
27:56Additional
27:57that's that's nice thing about uh this.
28:00I don't have to type right. additional
28:01analysis.
28:02>> That's actually my favorite feature
28:03about AI that I never have to like if
28:05I'm I make a typo, I never go back and
28:07fix it.
28:10>> Exactly. With the results uh using Let's
28:13just And we could use our favorite
28:14Amazon Bedrock, but let's use Claude
28:17Claude um
28:19>> SDK
28:19>> SDK
28:23uh Claude's AI SDK.
28:28All right. So, now it's going to kind of
28:30understand what we want and make some
28:32updates to our design.
28:34And it's actually even going back to our
28:37requirements and making some additional
28:40changes there. So, it's smart enough to
28:41know like, hey, I know you're updating
28:42the design, but the requirements aren't
28:44quite what you mentioned. Now, let's go
28:45and update the the requirements as well.
28:49So, now we have uh a microphone button
28:51that you can click on and now it it will
28:55transcribe it and it will keep going
28:59from there.
29:01There's also, I didn't mention this at
29:03the beginning, but
29:05we do have uh an idea of checkpointing.
29:07So, if you're going through this process
29:09and you were like, "Oh my goodness, the
29:11it's way off what I expected," you could
29:14just click this checkpoint button or in
29:16this case the restore button and it
29:18would restore back to its previous
29:20state. So, this means that you can
29:22easily go back and forth instead of
29:24having to rely on your version control
29:26system and revert the changes and go
29:28back
29:30Uh, so it also kind of makes little
29:33changes. It might find an error. It kind
29:34of goes back and forth and updates it.
29:36So now it's creating it's updating the
29:39design document based on our suggestions
29:42and we can see what it's doing. Yeah. So
29:45now here we have the transcribed text in
29:48our interface as we expect it would be
29:50there and some better analysis.
29:53It's also adding in.
29:54>> So what's the analysis? Go back.
29:56Analysis result. Okay,
29:58>> got it. So that's the that's what'll
29:59capture the
30:01the the results. There wasn't really a
30:04results interface before, was there?
30:07No. Mhm.
30:11Perfect.
30:13I love how happy the
30:16gentic ids are. Oh, perfect. I've
30:17updated the design document. So we can
30:20now take another look at it
30:24and we can say okay now we
30:26>> could add the transcription service to
30:27the to the system diagram too.
30:30>> Let's see. So we have start screen
30:32question result recording component
30:34local storage web audio ID. Oh says
30:36speech detects
30:37>> and it added the cloud AI too.
30:40>> Perfect.
30:41Perfect. All right.
30:42>> Honestly when when this system diagram
30:44looks good I'm like yeah looks good to
30:45me. Let's move forward. LGTM.
30:49All right, let's move to the next part.
30:52So now we have we have taken this idea
30:56that we had we we updated the
30:57requirements went into the design and
31:00now we are going to the last phase which
31:02is the task. Well, I guess it's more
31:04like the second to last because once we
31:06have list task we need to implement
31:08those tasks and there's ways we can do
31:10that as well.
31:12And this workflow is um there's a
31:15actually a lot of feedback we've gotten
31:17on this workflow on improvements we can
31:19make. But in general it's pretty
31:21flexible because it's all built in
31:22markdown. So you see there's like this
31:24refine button on the top. So what we
31:26also have a flow is if you go and like
31:28make a manual edit. You can refine that.
31:32Clicking refine will basically take that
31:33edit and and kind of introduce
31:36integrate it into the rest of the design
31:38doc. So if you do that in the
31:39requirements, it'll generate like the
31:40user story in acceptance criteria in
31:43years format which we'll talk about. Um
31:46and you can same thing on the design and
31:48actually even same thing on task list.
31:49So like let's say you've done three or
31:51four tasks and you're like hey I
31:53actually want to change something cuz
31:55often this is in real life right we
31:57implement four t four tasks and we're
31:59like hey actually like now I want to
32:00change something cuz it didn't this is
32:02not what I wanted to do. you can go and
32:03make changes out of band and and
32:05clicking the update tasks will refresh
32:08uh the will refresh. So Kira will look
32:11at your codebase and actually see what's
32:12done and and uh update the the upcoming
32:16tasks as well. So now it's at a part
32:19where yeah we have this task list of all
32:22these tasks uh and we also have an
32:24option to keep the optional tasks or
32:26make them all required. And here are
32:28these property based testing uh which
32:31are all optional right now. Let's what
32:34do you think Nikl should we just keep
32:36all optional tasks for fast?
32:38>> Yeah, let's just keep it optional test.
32:39This was actually a flow a lot of people
32:40asked us when we we didn't have this
32:41flow when we first launched but um what
32:44was happening was that for every feature
32:46you had a lot of unit tests and people
32:48sometimes wanted to see the see the end
32:50product before they get to the tests but
32:52they wanted the test. So we introduce
32:53this flow to keep tasks in this uh keep
32:56tasks spelled out but set to optional by
33:00default.
33:03>> Okay. So
33:03>> should we take a look at the task list?
33:05So we can do the same thing here. We can
33:06refine this task list as well now. So uh
33:09is it can we see like uh I keep
33:11scrolling down. So you got create the
33:13data models, implement audio recording
33:16service,
33:19implement response validation storage,
33:22the AI integration. One thing you may be
33:24thinking about, hey, I have this list of
33:26tasks, but I want to get to something
33:27like right away that I could test.
33:30So another idea too is we could refine
33:33this list and reorder it.
33:36>> Yeah, let let's ask I I I think so. So,
33:38I think let's let's ask Hero to
33:41like I'd say prioritize getting
33:44[clears throat] an MV how how should we
33:46frame this? Prioritize getting an MVP
33:49scope together by task five,
33:53right? So, we have something to see.
33:57There's a product there's like a product
33:59analogy here where like you when you're
34:00building a car you first start with a
34:02skateboard and move to a bike and then
34:04build a car or you can choose to build
34:06the wheels you can build the engine and
34:07then stitch together stitch together
34:08everything at the end. The better way to
34:10build a product is is you you build
34:12interim steps so you always have
34:15something that's workable,
34:18right? Yeah. This is very flexible. We
34:21can kind of move these tasks around. And
34:23I kind of like that idea of getting to
34:26the first five tasks that has an MVP we
34:29can show you, especially in this time
34:30short time we have together today.
34:33So that I think that that would work
34:35well. I also sometimes just have it
34:38condense tasks down. Like if I see it
34:40has like 25 tasks, I'll be like, "Whoa,
34:43why do we have 25 tasks? Can you just
34:45can you make it into like 10 tasks?" And
34:47then it'll kind of group some more
34:48together. or some people who does test
34:51first development who who writes their
34:53tests first.
34:55Okay. All right. A couple people over
34:57here.
34:58>> No one likes tests.
35:00>> So maybe that's the workflow that you
35:01like. Then you can set this up. Write
35:03all the tests first. Let's do test first
35:05driven development.
35:07All right. So we have now we have I
35:10think for the for this uh a core MVP
35:13with one, two, three, four, uh five
35:16tasks. even has task two or phase two.
35:19So let's see
35:20>> is there UI state captured in the first
35:23five
35:23>> setup dependencies create core data
35:25types implement application state
35:27management implement audio recording
35:28speech test implement response
35:30validation storage
35:33create start screen component. Oh, so it
35:35looks like the first phase may not even
35:37have the UI.
35:39>> Should we should we ask it to have some
35:42UI? Yeah. Adds
35:45uh UI to the initial phase
35:49>> so we can actually visually test it.
35:51>> So we can visually
35:54>> play with it. Yeah.
35:55>> Test it.
35:57>> Good call out.
36:00>> That's we did a run through of this.
36:02Like it's it's interesting because we
36:03did we've done like now three
36:04runthroughs of this and every time it's
36:07a little different, you know, like every
36:08time like last time we did it, it it it
36:10was it got it got this in the first run.
36:13So it's it's it's super interesting.
36:17Yes. That's why you always read it. You
36:19I mean we have to all when we're using
36:22coding agents and and Gentic IDs, you
36:24have to read it and pay attention to
36:26what it's doing because otherwise it
36:28will do things that we don't expect. All
36:30right. So, it looks like it it reordered
36:32it again. Now, it's it's task one
36:34through 10, though. Um, until it gets
36:38into phase two,
36:41I think for the sake of time, because we
36:44only have about 20 minutes together,
36:45let's let's see if we can start
36:48implementing some of these tasks.
36:51And I do have an after. So, if it
36:55doesn't get through all of them, we'll
36:56still show you what it comes up with
36:58afterwards. How's that sound? Yeah,
36:59sounds good to me.
37:01>> So, we'll start task here. And you can
37:03see we can do start a task a couple
37:04different ways. We can just click in the
37:06UI or we can talk to it in the chat. So,
37:09I'm going to start task. Start task.
37:12I'm just How about we do five?
37:15Yeah.
37:15>> And we'll do six because that has the
37:17start screen component.
37:18>> Yeah. And right now, what it's doing,
37:20it's it uh it's going to go into the
37:24terminal. It's going to run some
37:26commands. Now, typically, you can see
37:28here we have a whole task list. Uh
37:31there's task the cued. So, it's going to
37:33do task one and then it's going to do
37:35the rest of the task afterwards. These
37:37actually need to be done in order. Uh
37:39that's why it it does it this way. You
37:42can also just take a look at it. It's
37:44now starting to do it. It this one was
37:46pretty easy. Just had to like make a few
37:48files and install a few npm
37:50dependencies.
37:52And then I can kind of look at the next
37:53one and see it's starting task two. It's
37:55going through it pretty quick. This this
37:57will probably take a little bit of time.
38:00>> So while this is happening, why don't we
38:02uh pivot to talk about property property
38:04testing?
38:05>> Great.
38:06>> Um maybe switch to the design view.
38:09>> Yes. Cool. So I'll just give a little
38:11bit of an overview. So there's a reason
38:13why we went with in the requirements we
38:15went with this user story acceptance
38:17criteria format. And the acceptance
38:19criteria are structured in what's called
38:22years which is easy approach to
38:24requirement syntax. Believe that's
38:26that's it. Uh but what what what that
38:28means is that when you define something
38:30it says when you do this the system
38:32shall do y and that allows us to create
38:36uh essentially logical equivalence and
38:38do things like formal reasoning on the
38:41actual requirements. So that's really
38:43what a property is. Uh you know there's
38:45a definition of a property here. it's a
38:46characteristic or behavior that should
38:48hold true across all executions of a
38:50system. So what we do is we take the
38:53year's format requirements extract
38:54properties and then from those
38:56properties we actually there's a
38:58framework called fast check that allows
39:00you to actually test those properties
39:02and unlike unit testing that has like
39:05limited inputs uh you know the inputs
39:07that you provide to the system to test
39:08the output with uh property testing you
39:11it actually fuzzes a whole bunch of
39:13values across a range. So it generates
39:16more evidence that you've when you have
39:18a requirement the generated output maps
39:21to the requirement right so that we're
39:23thinking about this whole world of specs
39:24in a very first principles approach
39:26right like how do you how do you solve
39:28the problem in the AI world where you
39:30where the out generated code actually
39:33matches the user's intent and property
39:36based tests are like the first way we're
39:37going about it but we're there's a lot
39:39more we have planned
39:42>> yeah it's pretty powerful everything you
39:44and do with it. I'll I'll show you an
39:45example. I did this uh I'm a I'm a chess
39:48fan. I created a I vibe coded or spec
39:50coded, I guess you could say, a chess
39:52app. Uh and I was for me and my son and
39:57we were we wanted to make sure that it
39:59worked as it expected. So here is a list
40:02of property based tests for this other
40:04application that we created. And you can
40:06even see if you hover over it, it tells
40:08you what the test was for. This really
40:11does test those requirements. uh fast
40:13check is by the way if you're using
40:14Python or a different language there are
40:17other libraries that do that do does
40:19this property based testing so I can
40:22even take a look at it uh it says all
40:24property based tests passed for any
40:26board to stay in the AI gener move the
40:27move must be valid against the movement
40:29registry and so here's the test and you
40:32can see the uh for those of you are
40:35looking at imports from npm it's this FC
40:38from fast check
40:40and we write here we write our test we
40:44do an assert on it we have create board
40:46state and then one thing that I find
40:48really interesting is this numbum runs
40:50at the end so it's going to it's
40:52basically doing fuzzy testing it's
40:54making it changes random changes to the
40:57test file and it'll actually run over a
41:00hundred times to make sure it is as you
41:03expected and I will say adding these
41:06property based testing makes the output
41:08of your code at a much higher quality
41:11and oftent times less bugs.
41:15And I you can see here it just goes
41:17through a bunch of different tests to
41:19check to see if the move is valid.
41:22Where are we in our demo? So we have
41:26step task five. Looks like it's on task
41:29six now which will take uh a moment.
41:33>> So as the tasks get done, it also marks
41:35it off in the implementation plan. So
41:39if you take a look at this.
41:40>> So here's our job seekers app we've been
41:42creating. You can see we have nice
41:43little check marks next to everything
41:46and it's diligently working on task six
41:50which is this.
41:51>> What happened right now is one of my
41:52favorite features we launched recently
41:54like the diagnostics tool. So because we
41:57own the overall ID interface when there
41:59are problems like so if you go to the
42:01problems tab in in your um editor
42:04typically reports like your your lint
42:06linting issues and the agent is able to
42:08read it. So we used to have a lot of
42:11problems with specs like you'd have five
42:14tasks generated and you open problems
42:15tab you'd see like there's like six or
42:17seven issues to fix but like ever since
42:20we've had this added this diagnostics
42:21tool it's been really great.
42:23>> So it completed step six. I'm not sure
42:25if this it wanted me to go to 10. We
42:27could keep it going. Let me see. I'm
42:29just doing an npm rundev because I want
42:30to give everyone here time to ask us
42:32questions and to show our quick
42:34optimized demo at the end. So, if I open
42:37up 5173. Oh, looks like it didn't
42:40finish. It still has a hello world. I
42:41think it probably didn't link the
42:43component the start.
42:44>> Why don't you just go ahead and do do
42:46the other ones and you can while it's
42:47happening, you can switch to the
42:50finished app so people have time to ask
42:51questions.
42:52>> Yes. Okay. We're going to cue all these.
42:55Uh you can see here though,
42:58>> last time we did this in five tasks, it
43:00was done. So you can see it's starting
43:02to create the components that that we
43:04want. Um some use states. These are like
43:08hooks in React.
43:10Okay, let's take a look at our final
43:13app.
43:16I'm going to run dev on this. You see
43:18here, same principles as the app I'm
43:21showing you. We have our spec design
43:23requirement task
43:26and so this is what it looks like. Uh
43:28it's an interview prep practice. So we
43:30have behavioral question, technical
43:32questions and leadership questions. It's
43:33connected to uh a claude sonnet 3.7
43:38model I believe. So which one do you
43:40want to try?
43:41>> Um just pick pick one. Let's do let's do
43:44uh technical.
43:46>> Sure. And I'm going to allow this time.
43:49So explain a a complex technical concept
43:52from your field to someone.
43:54Hello, let me explain about websites.
43:58There's a server and there's a client
44:02and our website has to interact. It can
44:06run locally. We can have single page
44:08applications. We can have serverside
44:09rendered apps.
44:12>> This could keep going.
44:13>> As you can see, we've gotten pretty low
44:15scores in our every time we play to
44:17this. [laughter]
44:19Probably not going to pass the interview
44:21with that, but for the sake of this
44:23demo, we'll keep going. It's actually uh
44:25one thing it keeps the last answer from
44:28the last one we ran in here. It's just
44:30something we didn't change yet. Describe
44:32a challenging technical problem you
44:33solved recently. What was your approach?
44:35What was the outcome?
44:38>> I'd say this app, right?
44:40>> We learned about the web APIs. We
44:42learned to create a specriven
44:45development to create this application
44:47I'm using right now. I learned about a
44:50little bit about Vit and React and
44:52property based testing and I got to do a
44:55demo in front of 100 people.
45:00Okay.
45:03How do you stay current with new
45:04technologies industry trends? We should
45:06ask the audience this. I'm sure they
45:07have lots of interesting answers.
45:10I I attend reinvent amazing events,
45:14amazing talks. I also go and look at
45:18builder center that AWS uses to tell
45:23people about our latest tech. Uh I go to
45:27conferences. I look at Stack Overflow. I
45:32I go online to YouTube and watch a bunch
45:35of technical creators.
45:37Maybe I'll get a little higher score for
45:39that. So then I click finish, get score.
45:42And now it's taking all this. It has a
45:44very simple system prompt uh that just
45:47does an analysis on it. And it should go
45:51and grab the information. It sometimes
45:53takes 30 to 60 seconds.
45:57>> Nice. You got four out of 10 this time.
46:04And of course, the candidates
46:06demonstrate some relevant technical
46:07knowledge, engagement with the industry,
46:08but the response lacks structure,
46:09clarity, and depth. Um, so great
46:12interview prep that we could use and
46:14literally just took us not so long to
46:17put it together.
46:19And looks like it did actually complete
46:21this as well.
46:22>> It'd be nice to see something a little
46:23more than hello world after.
46:26[sighs and gasps]
46:27>> All right. Well, let's give some time.
46:30Um, let me switch back to Do we want Do
46:33you guys want to see what it it came up
46:34with after step 10 that we're not faking
46:36this?
46:37>> I do.
46:38>> Okay, let's see here. Okay, this is what
46:41it This is what it came up with. Start
46:43interview prep. Look, tell me about
46:46yourself and your background.
46:48Allow this time. Test test test. Hello.
46:50Hello. Hello.
46:53Okay, so it's duplicating it. Maybe we
46:56could go back and have it do something
46:58else. greatest strengths. This is my
47:00greatest strengths. Test, test, test.
47:02Okay, you get the idea. One thing I
47:04didn't do is I didn't put my anthropic
47:05claude key in here, so it's definitely
47:07going to fail at the last step.
47:09>> Oh, you Oh, yeah, you forgot about that.
47:11[laughter]
47:12>> Okay.
47:12>> Uh, test test. And I think
47:17right here I maybe it's not even using.
47:20That's amazing. [laughter] That's all
47:22you need to do, guys.
47:27All right, let's switch our point.
47:30>> Well, I mean, I think we're we have 12
47:32minutes left. I I think we we got
47:34through the the core workflow that we
47:36wanted to talk about. I think just one
47:38thing to keep in mind in specs, like we
47:39launched this and like there was a lot
47:40of positive momentum and adoption around
47:43specs, but this is just the start. As I
47:45said, we're we're solving like first
47:47principle problems on the requirement
47:48side. How do you know often what we do
47:50is like we do things like scenario
47:52planning, right? like or scenar but like
47:55when we actually build something like
47:56hey when when you do when the user does
47:58X and Y what's the expected behavior
48:00right that doesn't really translate well
48:02into like the structured kind of doc we
48:04have here but there's clarity on
48:05requirements as a concept how how do you
48:08know your task list matches um uh your
48:11task output matches your intent is
48:13another area we're just continuously
48:15going to invest in and I'm actually
48:17super excited about this journey with
48:18specs we're on and I'm very excited to
48:20come back next year and talk about all
48:22the cool things we do with specs over
48:23the cold over the year.
48:25>> Me too. And also just looking at it, it
48:27never implemented the cloud AI
48:28integration. That was the MVP didn't
48:30have that integration,
48:31>> but you got a 9 on 7 on 10. So, you
48:33know, it's better off.
48:34>> I know. [laughter]
48:39>> A git repo. Um,
48:41>> yeah, we'll put it up. Sure,
48:43>> we can put it up.
48:44>> I don't have one right now, but we we
48:46can certainly do that.
48:48>> You could use the oneclick git
48:50integration on Kira. Oh, really?
48:52>> Yeah,
48:53>> that's We'll have to try that.
48:56>> You just You click here
48:58>> and you add them to your source changes.
49:00Oh, you're you're back on Oh, you're no
49:02longer on the demo. Never mind. I was
49:04going to demo another feature. Cool. Um,
49:08>> thank you.
49:09>> Yeah, thank you. Any other questions?

About this transcript

This page contains the full transcript of AWS re:Invent 2025 - Spec-driven development with Kiro (DEV314) by AWS Events, generated from the public captions YouTube serves with the video. The transcript has 8,813 words across 1,288 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.