Getting more out of every Kiro credit

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Getting more out of every Kiro credit · Channel: Kiro · 58:44 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, Model inference and optimization

Summary

Kiro host Eric Hanchad and developer advocate Laura discuss ways to make Kiro credits go further, especially on the free tier. Laura shares strategies such as choosing less expensive models for simpler tasks, setting steering instructions, using spec-driven planning, automating work with hooks, making easy edits directly, and providing agents with clear context. They also demonstrate Kiro Crew’s persistent sessions and Discord integration, and point to an open-source benchmark comparing agent costs. The show also covers recent Kiro updates, audience questions, and a brief example of building a Fire TV app with Vega OS.

Key points

The free tier’s 50 credits can support meaningful projects, depending on the models and workflows used.
Choose models according to task complexity; simpler work may not require a frontier model.
Kiro’s Auto model selection routes based on task complexity, while keeping a consistent 1x credit multiplier.
Steering files can prevent unnecessary output, such as tests, boilerplate, and summary documents.
Combine related changes into a well-scoped prompt when practical to reduce repeated instructions.
Use spec-driven development for complex work to plan before implementation and avoid wasted changes.
Hooks can automate tasks such as linting and formatting; perform simple edits directly when that is faster.
Give the agent useful context up front, including relevant files, folders, images, or sketches.
Kiro Crew offers persistent sessions, concurrent agent tasks, and integrations such as Discord, complementing the IDE and CLI.
The hosts recommend experimenting with models and consulting the open-source sample agent cost benchmark when comparing costs.

Chapters

Start	Chapter	Summary
00:02	Introduction and Kiro updates	Eric welcomes viewers, previews the discussion about credits, and reviews recent Kiro releases and features.
07:04	CLI version 3 and upcoming events	Eric explains why to try CLI V3, mentions upcoming events, and brings developer advocate Laura onto the show.
10:53	Enterprise model availability and the free-tier article	Laura and Eric discuss company controls over model access and introduce Laura’s article on making free-tier credits last.
14:33	Laura’s project and choosing models	Laura describes creating a video reel with Hyperframes and explains why less expensive models can handle many tasks.
20:31	Auto routing and Kiro Crew	The hosts discuss Kiro’s Auto model option, then explain Kiro Crew’s persistent sessions, concurrent work, and Discord integration.
31:23	Learning Kiro and credit-saving strategies	After discussing ways to learn Kiro, Laura outlines strategies including steering files, combining related prompts, and planning with specs.
39:34	Hooks, direct edits, and added context	Laura and Eric cover automations, handling simple code changes directly, and supplying agents with relevant files or visual references.
46:28	Comparing model costs in practice	Laura demonstrates the different credit use of models on the same simple prompt and recommends matching model choice to task complexity.
49:42	Spec modes and cost benchmarking	The hosts discuss Quick Spec and an open-source benchmark for comparing coding-agent costs across providers.
53:30	CLI shell troubleshooting and audience questions	Eric and Laura answer questions about shell profiles, wrappers, and Kiro powers, including a troubleshooting note for Zsh users.
56:09	Vega OS demo and sign-off	Laura briefly shows a Fire TV app experiment using Vega OS before the hosts thank viewers and close the stream.

Quotes

I don’t need to use that same model to review the code it generated. — 19:58

you don’t have to pick one over the other. — 29:47

keep it simple and just, you know, set up some steering files — 36:51

do the easy stuff yourself if you’re able to. — 42:13

Concepts and entities

Kiro, Kiro Crew, Kiro CLI, Kiro IDE, Laura, Eric Hanchad, HyperFrames, DeepSeek V3.2, Sonnet 4.5, hooks, steering files, Spec-driven development, sample agent cost bench, Vega OS, AWS

Full transcript (original language)

00:02 [music] [music] [music]

00:37 Hey everyone, my name is Eric Hanchad and this is the Kuro weekly show where we talk about all things Kirao. This is our live stream. I I love our intro. I think we have one of the best intros. I’m always it just it just pops when I I listen to it. Uh, so if you’re here watching right now, we have a chat room. So jump into the chat. You can, we’re streaming on LinkedIn, YouTube, and Twitch right now. So any of those places, send us comments. Let us know.

01:08 We love to stop the live stream and answer questions right in the middle at any time. So please, yes, ask us any questions. Also, let us know where you’re where you’re watching from today. So, I do have a special guest today. We’re going to talk a little bit about Curo credits and and how to save on your Curo credits. I think that’s a big topic. We we keep hearing it over and over again. But the way I like to start every one of my shows is to check out our curo.dev. So, let’s bring that up.

01:42 Here is uh curo.dev. So, uh what I like to do is we’ll go into our resources and look at our change log. So, this is the place to be to first, if you are brand new to Kira, have never heard of it before, this is the place to go. But also, you can take a look at the latest updates and I’m hoping some of you who are watching today are Kira fans and want to know what’s new to Curo. So, just today, we just launched workflows, safer, untrusted workspaces, enterprise

02:12 signin controls. So, you can run reusable multi-agent workflows in the background, apply stronger safeguards. This applies for the IDE 1.2 introduces workflows. Um, you can run the multi- aent workflows in the background. I’m It’s kind of hard to describe workflows without like showing you. So maybe later in the in the episode we’ll show you. Uh, we’re also creating some content on this to kind of give you a little bit of more of an idea of how how it works. Um,

02:44 but you can ask Hero to run a usable multi-step plan while you continue in the parent chat. Each step runs in its own agent session. Use the workflows area above the chat to inspect progress. Open step. So, it’s almost like your uh you can give it like these these workflows, these steps, and it’ll like uh autonomously work on each one of them and they’re reusable. Uh it’s very little little difficult to to explain, but it’s a nice new feature that we’re adding in to the ID. It’s already in Curo Crew. So if you are a Curo crew

03:16 fan, you can use workflows uh in in there today. And looks like yeah, we added in this enterprise signin control. Administrators can restrict which signin methods are permitted, prefill your organization signin deals, add help link. So if you are a Curo for enterprise customer, we’re always adding new features there. So you can check that out. Oh, looks like uh Hannah, hey, thanks for joining us from Chile. I appreciate it. Agent focus mode, session details,

03:49 regional data handling. Awesome. And if you didn’t see, I think our last live stream, uh I’m just checking my calendar, was on the 23rd, but since then, we had a very big release. One one of the questions we often get is when is or what when are you getting X model and and one of the latest models that we were happy to introduce is cloud opus 5.5. It was released September 28th. So it brings stronger gender coding. I love

04:20 Opus 5.5. I have been using it so much recently on so many projects. This weekend I just all my favorite projects I’ve been working on been wanting to work on. And I said, you know what? I’m going to use this for 5.5 and see how it does. It did amazing. I’m going to show if we have time later on, I’ll show you a a fun little project I created from it, an MP3 music player. But you can see, yeah, it’s available now. Uh, it’s available with experimental support. It just keep an idea, keep this in mind that it’s for Cure ID, CL, CLI,

04:54 Crew, all the different services. But you do need to be a paid plan uh on our Pro, Pro Plus, Promax, Empower customers, and US East one, and US Central. So, you definitely need to do that. Also includes a 1 million to token context window. Hey, a couple more people are coming in. Yeah, if you are watching live right now, let me know where you’re watching live from. Uh in chat, love to answer questions in chat. Love to see where you’re all from. Hi, Devx Cuba. From Devx Cuba is from Cuba. Who who knew?

05:26 Who knew? Uh Stephen from Ghana, thank you for joining in. Azie from South Africa, thank you all for joining in. This is the Kira show. Also, power management v3 output styles, full screen scroll speed. Yeah, some nice new updates for the CLI. So, if you’re using Kira CLI, continuously add some neat new updates there. You can now install and uninstall powers by using the slashpowers in B3. By the way, if you have have if you’re

05:56 using Kira CLI, make sure you use the d-v3 when you start it. That’ll get you version three of it. Look, we got someone from the US Baltimore. Hey, hey, Thomas, thank you for joining us today. And now there’s an there’s a slash settings and a V3 settings. There’s a settings display to select full screen and scroll speed. Lots of really nice terminal UI improvements to make it even better. Run a hook when a V3 session ends. So now we added uh if you haven’t used hooks in Curo, they’re a great way to like trigger something when something

06:30 happens. And now there’s a new hook to run a hook when a V3 session ends. Okay, we got our LinkedIn user doesn’t say your name, but thank you from Germany. Uh yeah, so lots of cool CLI updates, workflows, and then uh even configurable terminal timeouts, clickable file links for the ID. So lots of new improvements for the ID as well. Uh we do have a guest. I’ll bring it bring her on in a minute. Oh, Alp has a question. Why we should use V3 and CLI? What’s the

07:04 advantages and differences? So B3 is the brand new version of the CLI. It’s kind of been in like experimental mode or preview mode for a while. Soon it will be the default mode in CLI, but it has all the brand new features. It has a different terminal UI look and feel to it. So that’s definitely where we’re headed. So when you use uh when you use the CLI, Cur CLI, use the V3 version. Like I said, you can use d-b3 with it.

07:38 Let’s take a look and community. Uh we’ll look at upcoming events. No. Uh no new upcoming events if you watched last week, but the cutscy Montreal the Cusack. I will be in San Francisco. It’s like in just a couple of weeks in October 15th for Versel’s conference. From there we will uh it’s mostly going to we have a booth at that verse conference. I will personally be there. So if anybody wants to talk about Kira,

08:08 I’m happy to talk to you guys about Kira or if there’s uh I I’ll be around. I’ll be around. Lot lots of fun things happening happening. There’s also some other events that Kira’s going to be at. I don’t think they’re in the event calendar yet. We’ll be updating it soon, especially in November for a couple of AI conferences, which I’m excited about. All right. Oh, we got another Hello from Mexico, LinkedIn user, Mexico. Hola, come on in. Thanks everybody for joining and watching today. So, uh Oh, I I see a question, but I don’t see an an arand,

08:46 I don’t think your question came through, but please ask a question away. But let’s bring on our guest today. Hey Laura, how’s it going? >> Hey everyone. Hi Eric. How’s it going? >> Hey, great. Great. So, Laura works uh on on the same team I do. She is a senior developer advocate. How you doing, Laura? Can you introduce yourself to everybody? >> Yeah, definitely. Hi, everyone. Uh Laura here and I’m also a developer advocate for AWS. Um I’m based out of Seattle. Um, so

09:18 that’s where I’m coming to you live from. >> Awesome. And one of the reasons uh we wanted you on today because uh we by the way Laura is great. She lots of the things lots of the feedback we’re hearing from the community. She has been like leading up that feedback loop trying to make sure that everything that we’re hearing is getting to the right people. uh we’re we’re working together on a lot of different stuff on the Kira world, especially on content. So, it’s great to

09:50 have her on this podcast. And she was on this live stream. You were just you did a live stream last week, right? >> Yeah, we were just on a live stream for the uh AWS YouTube channel talking again about some um Kuro features and uh I think that actually just went live um on demand. So if you are looking to learn uh more about curo and building a game and spectrum in development um you can find that at the AWS uh YouTube channel. I think it’s called cloud report is the name of the live stream. >> I love it. And what and I always ask all

10:22 my guests what are you drinking today? I I have I’ll start uh as always I have my my protein shake my berry protein shake uh that I’ve been drinking every morning. What what about you? Um, I I just have water in a Stanley cup, so nothing too fancy, but I will be taking sips. >> Uh, I love it. Yep. The I think the winner of most my guess has been water. I know this is a little bit early, too. So, >> coffee maybe. Yeah. >> Yeah, coffee. All right.

10:53 >> Coffee. >> Uh, Brendan did have a question. Okay. My company uses Kuro, but the models available are limited. Is this company Is this company managed? Are these models based on the offer that Amazon Bedrock has at the enterprise level? My bedrock shift enter. I think you got cut off there at the end, but I think I get the gist of the question. What do you think, Laura? >> Um I I mean it depends. Yes, I think there are um enterprise controls that your company can um kind of turn on and

11:24 off certain models for you. um probably for you know pricing purposes to make sure that um you know folks aren’t maybe racking up too large of a bill I imagine um but yeah I mean underneath the hood the Kuro um harness is using uh the models that are supported by bedrock so they’ll be supported by bedrock first and then you’ll see them kind of propagate over to Kira um what was the other aspect of the question I think it was like are they limited by the company I think they can be but um they don’t not by default I don’t believe unless

11:55 the company does like selects that in one of the settings. >> Yeah. Yeah. I think you’re you’re ex exactly right. It’s unless the company decided they didn’t want you to have all the latest models. Uh you should you should have all of them. Yeah. But they there’s a lot of enterprise controls that I don’t even know about. They they’re adding more all the time just to make it so it’s very fine grained. Uh oh, we got a Hi, Laura. Thank you. >> Hello. >> And hi live from Mexico. Marco, thank you. Uh yeah, let me know what you’re drinking in the chat. I I see some

12:27 coffee. Got some coffee users here. Um he says at the rate uh burning tokens makes sense. [laughter] Well, you know, we have the right person on right right now today because uh Laura has been top of mind. Laura wrote a really cool blog post about token usage and pushing Kira’s free tier to its limit. Let me I’ll I’ll introduce it, but maybe uh I know you have some stuff you want to share, too. >> Yeah, >> Laura, let me see here. This is kind of

12:58 this is a very fun article. Let me throw it in chat right now. Um so, this is uh this is an article about like where how you deal with the free tier. If you know, if you go into events here and go go into our cure.dev to have pricing. You can see here that we give you uh we have all these different free plans with the $50 credits for just to try it out. And I would say like the 50 credits is just enough to maybe just get just a little bit of of work done.

13:31 Laura, would you agree? [snorts] >> Yeah, I mean I think in the blog um and I could talk about it a little bit after. Uh I got a a decent project um two decent projects um small but but nothing um insignificant either out of my 50 credits. Um and again I think the point is like depending on which model you’re using um the rate at which you consume those credits um can vary. Um and that’s I think what we’re going to talk a little bit more about today. But, uh, yeah, I I I was able to stand up, um, my personal website [snorts] on it.

14:01 And then I was able to do, um, kind of a quick, um, marketing like real video, um, with this open source package that I’ve been playing around with a lot, um, the last couple months. Uh, and both of those consumed I think I read it in the blog, but I think the the video real project consumed about a little under 18 credits. >> Let me bring up your your screen then. Let me see here. Sure. >> Uh let’s see. Let me take mine down and yours up. Okay. >> Am I up? Okay.

14:33 >> Yep. >> Yeah. So, in the same blog that um you you all have in chat um down kind of a towards the end of the blog, I talk a little bit more about the specific project that I built and and you know, which models I was using, how many credits I consumed. I even share the prompt that I used for that specific project. Um, and the idea here is there is a really cool um, open- source project called Hyperframes um, which I think has been making making a little bit of waves um, not just around my

15:03 team, but I’ve heard a lot of other folks use it. But it basically um is a a framework for um, agents to write HTML and then render video out of it. And a lot of the um developer advocates on our team have been playing around with this and taking a look at like building B-roll for a video uh and building uh like teaser reels for when we’re at conference floors and conference expose to kind of have like a video playing in the background as we talk to you all about uh whatever it is we’re there to talk about. So anyway, this was the uh

15:37 project that I was working on and what I based part of the blog on. Um but I was creating a 30 secondond clip for Kira itself. Funny enough, so it was a very meta project. I was using Ko to build a video about KO. Uh, but I I was talking here about what prompt I was using, the two models I was using, and I was keeping this all in the free tier for the purposes of the blog. So the models that I used here were Deepseek u at the time V3.2 and then Sonnet 4.5. Um, and I was giving some snippets here about like, okay, what does this look like for

16:10 use in the um Curo IDE? So, here was a snapshot of like the first uh response to my prompt, which was about a little under nine credits. And then I showed here like what the actual uh reel that this uh 30 second um package generated for me. So, you can you can see that in the blog there. Um but yeah, I think it’s important to talk about uh how to how to maximize um if you’re someone who is exploring um different free tiers from from various vendors like I think it’s important to try to get the most out of you know the the 50 credits uh in

16:44 this case for Kira. Um so I write here and I think it’s easier to kind of talk about like >> can you bump up the pawn a little bit? >> Oh, sure. Yeah. Is that better? >> Yeah, >> a little bit more. Um but yeah, anyway, I have some I think it’s four or five, I have six. I I kind of wrote about like these six different strategies that I like to think about when I’m working with um in this case a free tier where I’m really constrained on um how many uh credits I have to build. >> Before you get into that, Laura, I want to ask you this. So, this is really

17:14 cool. So, >> uh you’re saying that you with the 50 credits because a lot of people see the 50 credits and they think, well, I’m going to use Opus 5. Well, I’m going to use You can’t use Opus 5.5 in the free, but I’m going to use the like the highest multiplier free one I I can get my hands on. >> And then I’m just going and then that’s going to just burn through all your credits in like, you know, an hour, hour or two. But you’re saying like you don’t have to use those latest state-of-the-art models that are like 2x, 3x. When you open up Curo, you can

17:49 use Deep Seek and you can get a lot of work done. You can use miniax uh some of the clawsonic ones that are even less than 1x multipliers and you can get a lot done. >> Yeah, exactly. I think I also included a table here and this was uh this is a little dated so some of the models here might have changed on the on the IDE um in terms of like how much um maybe there are better versions now of these. But anyway, my idea here was like let’s take a look at the models that could be workh horses in their own league. Um, and so I

18:22 outlined here, uh, the cheapest option, um, at the time, maybe it still is, was Quint Coder Next, um, which was, um, in some benchmarking was still showing some really good, um, you know, capabilities for code generation. Um, and then there are some other, uh, items that I noted here is like, okay, well, what are some budget friendly kind of in between options before you actually get to the frontier models? Uh, and so I indicated here like Miniax 2.1, which there’s a later version now. I think it’s 2.5. Um, but yeah, same with DeepSeek, right?

18:55 Like the idea being that you don’t have to always use the Frontier models to get a lot of the same functionality. Um, but you will be paying upfront for it. You know, in this case, it’s like almost two times as much depending on which model you’re using. Um so you know some of us who maybe are using um our identic idees or you know CLI etc identic AI in general from maybe a company perspective maybe you you don’t feel like you have to think about uh how much you’re burning through but I think a lot of us um are also building projects on the

19:27 side or I’ve talked to a lot of students as well who who obviously don’t have that flexibility so thinking about how to make those credits go further um whil still building uh you know good projects uh is is important. So, so yeah, to your point, Eric, like I think my approach is I don’t always go for the top tier model first unless I know I’m building something that requires like that that highlevel um reasoning from a model if I’m just, you know, doing a simple code

19:58 review of another model’s uh code. I don’t need to use that same model to review the code it generated. And actually that’s usually another um good practice is if you are generating code and you want to do a review, you would use a different model to review code that’s being generated by another model so that there’s slightly less bias there. Um so >> oh uh good question in the chat here. Wouldn’t a configurable routing system inside KO itself would be super helpful and enable us to bend our credits based on our preference based >> well funny you mention Arandan um I

20:31 think if you can see my screen here uh I’ve just switched uh Windows to the IDE uh and in this little dropdown let me I can bump this up a little bit um in this little drop down where you can select model we do have um the Kuro harness have that option it’s called auto um which does like it says here it chooses the model by task based on uh how complex your prompt is. Um so we do have that capability already. Um but I think in testing I think most of the auto um

21:03 routing is usually going to like one of the sonnet models or sometimes some of the smaller models but you don’t you don’t know exactly which model it’s using. So, you know, you kind of are getting this like one credit uh multiplier versus if you wanted to be really particular about how much you’re spending. You could even go down to the uh you know 0.5 multiplier of a model um to try some of that. Same same for some of these other models. So, that does exist. Um but it’s not to say that you couldn’t uh you know bump that down even

21:35 lower. >> Yeah, I think I think he was saying I think absolutely automotive is amazing. It’s glad it’s already there. You you could do this manually too, I suppose. Laura, right? Like you could like go to the plan mode and then switch it to opus and then go to the code and switch it to Quinn. One thing I noticed is like if you use one of the the the cheaper minax deepsee models and you use something like Spectrum in development mode, it actually does pretty good like when you get to that implementation phase because it has all the data in there. You don’t need to have the state-of-the-art model

22:07 all the time if you have like very detailed plans, a detailed requirements document, a detailed >> design document. Um, it can work well and you could experiment it with it too. Uh, I guess here’s a follow-up to Rendan, but instead of relying on the model discipline to follow the steering files, would be nice to have a deterministic approach from the harness itself. >> Yeah, I think I think Laura kind of answered that already with the auto mode. Um, especially because the because in the model auto we have no idea where all the credits go. But it is it doesn’t change the multiplier in the auto mode. >> It’s not going to like

22:40 >> randomly change it to like 2x because it decided in the background to use a frontier model and then it goes back to.1x. It’s just kind of that constant 1x. >> Yeah. Um, I think the other thing I was getting to, unless there’s any other questions, was um, >> let me answer one more before you jump in there. Um, >> when are we getting GPT61? I That’s a good question. We, uh, we are always adding in more models all the time. I wish I had like a great answer for you, but that is definitely top of

23:16 mind to get more uh, models in the new GPT6 ones, um, Astra ones in there. For sure. For sure. Uh I really like I really only ever leave Curioata. Perfect, Robert. Yep. Exactly. It’s it it’s surprisingly it will use like the latest models for a lot of things. So it it’s well worth using. Um but we have no way to configure that to our liking. Yeah, I mean Brenda, I get I get what you’re saying. Definitely

23:47 something we can think about for feedback. Um >> it’s worth bringing up too. I mean, I understand what you’re saying. Like the idea being that even if uh there is some preference by the harness on which model it’s picking in auto, it would be maybe it would be nice to be able to expose that to the user like, hey, use auto, but maybe only use these these certain models within auto. I don’t know. It’s something worth bringing up. Yeah. >> And Van Dread, no, that’s what we’re saying. Like you’d be surprised. What Laura is saying right now is you can do

24:20 so much with the 50 credits on the free plan. Uh that’s what this article is all about. Like I know you said 40 credits, but there’s a lot you can do for sure. Um oh, here’s one more question before jumping the the article. I want to like Kira Crew, but I don’t know what I’d use it for. Why would I use Kira Crew rather than Kirao? I feel like uh so there’s this idea of like personal assistance out there and it’s almost like a developer personal assistance is cur for developers. So you

24:52 have uh it’s kind of analogous to to Hermes or or openclaw but you can uh schedule tasks, you can have memory built in for it. Uh you can kind of work on the go. You can put it on your own Mac Mini if you wanted to or I have it on running on my own virtual private server. And then I have tail scale and I’m like VPNing in from my phone and doing tasks while I’m out running and I’ll be like, “Okay, can you do this?” And it’ll research things for me. So I I do it for like it’s like the perfect like developer personal assistant for

25:25 me. But you can also use it just like replacing the the IDE or replacing the CLI. It’s almost like you take what the CLI has uh and then you have a visual graphic interface for it because the CLI is just a prompt window. It’s just like a chat box has a lot of features um but you’re kind of chatting for and that’s like the main thing with it. Uh now you have like this whole system built around curo crew which is this whole app. It’s visual. You can have multiple sessions running at the same time. It just added in workflow as we mentioned earlier. Uh

25:58 and then you can have multiple like autonomous agents running at the same time. You can have it running for longunning tasks. That’s what I’ve been using it for. Laura, do you have like a more succinct way to put it than I? >> No, I think that’s that’s what I’ve been noticing, too. Um, funny enough, I I had a similar question myself when Kira Crew came out, and uh I actually been using it a little bit more over the last couple weeks, and I agree. I think one of the biggest um differences for me is the persistent memory, which uh you know, Kira right now doesn’t actually have um out of the box. It’s coming. Um,

26:33 but I think that’s one of the things right now that is actually really interesting about Kira Crew is you can have those sessions like Eric mentioned. And actually, I can probably even bring it up. Let’s see if I can. >> I really like too how it has so many interfaces in there. Like you can connect it up to Discord, you can connect it up to WeChat, you can connect it up to a whole bunch of things. >> Yeah. So, so I have it here and I think um it’s I think with Kuraku it’s easier to show and to see it than to talk about it. Let me bump up my visuals here a little bit. But yeah, so on the left

27:07 hand side um like Garrick was mentioning you have all the sessions in one place so you can see them all kind of at the same time. Right? If you were just using the CLI you’d have a bunch of different tabs or you’d have a bunch of different terminal windows open which can get unruly. Um here each session from within the Kira crew setup here is basically sitting on top of the CLI. So I don’t think we mentioned it at the top but Kira crew runs um over agent context protocol or ACP on top of the cure CLI. So it’s still using it’s basically Kira

27:39 as uh the the agent base for how all this is running. Um, but anyway, so yeah, all these sessions here are basically instances of the CLI, but you can see them all in one place and you can continue running tasks concurrently across different sessions. Um, and I think what Eric was mentioning, which I I think I found really valuable, is uh, you know, if I close crew or if I’m moving around and and not necessarily at my laptop, the session state persists. So if I, you know, were to do this on the terminal, if I close the window or

28:11 if I close my laptop, like the session’s done and then I have to open up a new session and, you know, talk to the agent again about what I was doing, which can be really annoying. Um, and yes, there are ways to kind of continue the session by doing like a SL session on the CLI, but you know, you have to kind of think about it and work at it versus it’s just always here. Um, so I think that’s a really cool difference about Kira Crew. And yeah, I’ve also been doing some like on the-go kind of stuff. So, this session here that I was playing around with was um you can see here, it’s a little hard to see maybe in the chat,

28:42 but uh I can control it via Discord. So, you see here this like little um >> header here that says like driven from Discord. Um I can interact with this particular agent session from in this case my phone. And this was a sample that I was playing around with at a conference last week where I was on stage and was interacting with the session from my phone and then uh you know asked the agent to like give me um a status report on the repo that I was working on. So that’s that’s really cool. Um and it’s really easy to set up

29:14 from crew. Um, not to say that there aren’t ways you could do this with Kuro, but like it would require some MCP servers and like just more complexity versus here it’s like a couple clicks on the settings and then you have it set up. So, I think those are my two um big drivers for why Kuro crew over Kira. But I think the whole point is like you don’t have to pick one over the other. Uh I think one major difference that you can just see from this UI is there’s not really a concept of like uh you know project directories, file trees like I

29:47 can’t edit directly into an editor here. There is no editor here um in the way that it exists uh you know in the IDE. Um, so you know, if you’re someone who wants to be like really deep into the code and work through your your project directories and you know, kind of have this VS Code type of view, like use the IDE and then if you want to be maybe on the go or doing some precision state stuff, um, switch in between. Um, but I think that the answer to the question is you don’t have to pick one over the other. It’s kind of like a combination

30:18 of the two. And underneath the hood, it’s using your same KO plan, so you’re not like having to open up a different account or something. It’s it’s the same account. Um, so it’s using the same credits underneath the hood. >> Yeah, well said, Laura. And I’m going to steal that demo. I I did a Kuro Crew demo for a local meetup here and I just I did some slides and then I just showed Kuro Crew, but I’m like I should have I should have thought of like connecting it up to Discord and be like, that’s a really cool demo that you could be on the go and >> talk back. You could even have a whole bot. I guess you could have like a Discord bot that powered by Curo Crew

30:53 that I don’t know talks to people and sends updates to the channel. Like I know streamers use this all the time when they go live. They’ll have like the their bot go in and update it. But that No, that’s a really cool use case. Uh one last question. In my experience, using Cloud Opus 5 for an analyst analysis tasks have saved me credits in the long run. Paradoxically, is it just me? Oh, wonder what kind of analysis tax

31:23 tasks are using LinkedIn user. Um, I’d have to find out kind of what tasks you’re using it for, but definitely um I would experiment with different models to see which ones work best for you. There’s there’s almost like this mythos out there that some people think that like claude opus models are better for for different types of design and maybe use GPT for uh specs or requirements or there’s there’s different philosophies.

31:56 I feel like they’re very similar as long as using the latest front of the tier models. like I I it’s hard to tell the difference of the outputs, but you might be getting a little little bit better output when you’re using the for certain analysis. I don’t know. Uh what’s the best way to learn Curo? Uh I know what do you think, Laura? >> Um I think there’s a variety of different ways. I think from our perspective, like we do have a Curo channel on YouTube. So cure.dev dev um do tdev

32:27 um uh is kind of a great place to go if you’re uh someone who consumes um learnings via video. If you’re like a visual learner, um we have great videos there. Uh we also have uh like Eric mentioned, there’s blog posts that um we’re kind of consistently writing and kind of teaching people through um written content as well. Um what else? I think those are the main kind of two vehicles that we have. Um, >> yeah. Our YouTube channel. Did you mention that? Yeah. >> Yeah, the YouTube channel.

32:58 >> Yeah. Yeah. Um, I know personally my per our personal channels, me and Laura write about Curo. We create content in Kira all the time. Definitely check out uh your what’s your dev 2 username? >> Uh, I think it’s L A U S A L I N. >> Yes. >> Yeah. >> We can put that in chat, too, because we’re both creating blog posts and articles every week almost on it. I have a personal YouTube channel. I just did a video today on about front-end development and using Kira Crew. Uh and and a little bit about a little spicy take on on how I think front-end

33:31 development is changing and it’s uh and it’s uh cooked. I think I even put in the [snorts] title. Um so it was a fun video that I did. to not to get too off track, but it was a little bit about how, you know, front-end developers are using these tools and these LLMs are there, but we still the taste that we have, being able to to see what a good design is versus a bad design. Those are really extremely important things. Um, being able to review code. So, we’re always always

34:06 going to need developers. Okay, let’s get back to the article. Um, go ahead. >> Yeah. Um, so I guess to recap for folks who maybe have joined a couple minutes late, uh, we were talking about this article that I wrote on Dev2, which is uh, pushing Kira’s free tier to its limits. We’ve talked a little bit about um, using different models that cost uh, less to drive uh, smaller use of your 50 credits a month if you’re just using the free tier. Um and then I was just about to talk about uh you know what are some

34:37 of the cost optimization strategies that I think about before I work on a project. Um specifically with the focus in mind that you know we’re trying to keep this to 50 credits to stay on the free tier. Um but obviously these same principles apply if you are not on the free tier and just want to make your credits from different pay tiers uh go a little further. Um so I think I have six different things I talk about here. The first is um we you know it’s not it’s talked about a little bit here and there but it’s I don’t think talked about uh as often as it should be but with a lot of these agentic AIS like you do have

35:10 the concept of steering files which let you uh basically inject instructions into um the model before it actually does anything with your prompt. Uh and so I think in this case there’s a lot of things that I see agents um doing quite a bit which in my case I don’t often need uh like property test or unit test when I’m just running simple demos. So I think uh a lot of my steering files are sometimes like hey skip writing um some of the testing and let’s just try to get a skeleton file going um so that I can try to make sure that the uh like demo

35:46 aspect works before we actually are pushing anything up to a repository. Um, and a lot of the times I found that when I’m working with different models, they, which it makes sense, right? They’re, they’re using best software development practices, um, they steer heavily to write these like really complex, uh, unit test, property tests, um, that just consume a lot of credits, um, when that’s not something that I necessarily need. Um so in these steering files the point is things that I consider worth putting in there um is something like

36:19 you know use minimal code like don’t do this whole boilerplate aspect um of writing code because it’ll just generate a lot of files and a lot of um it’ll spend a lot of tokens basically doing things that you don’t need it to do. Um, this is one that I use often is don’t add any tests unless you’re asked. Um, so that it doesn’t create any of these uh extra files that are just going to consume up my credits. Um, I don’t have it create summary markdown files unless asked as well. So a lot of the times if I’m just u maybe vibe coding and I don’t

36:51 actually need a full like project his you know history out of what I’m working on uh I put this into my steering file so that again I’m not generating files that I don’t want to at the end of the day all those files that are being generated are costing you. So, in my opinion, keep it simple and just, you know, set up some steering files so that the agent already knows off the top what it should uh be doing versus not doing. Um, for Curo, uh, I think this is still true, but it doesn’t cost uh any credits to set steering files up. Um, and it

37:23 just saves you from repeating yourself. So, that’s kind of what I mentioned here, and I’ll show you some steering files in the IDE in a second here, but that’s one thing. Um, this one I sometimes do. It kind of depends what project you’re working on, but uh if you have uh a project that you’re kind of scaffolding across uh like a bunch of different features, if you have a prompt that says like add feature A, B, and C, um all of that is kind of encapsulated in one prompt. So, it’ll consume like that many credits for that one prompt

37:56 versus if you have three different prompts going in to the agent, like each of those prompts will cost different credits. Um, so I think like I mentioned here, if the features or changes are related, uh, when possible, you can keep that in one prompt and then that way you’re also optimizing because you’re not sending in a bunch of different prompts that are kind of essentially saying the same thing or working on very similar aspects of the project or application. Um, so that’s another thing you can think about. Um, specs is interesting. Uh, so for those of you who don’t know, specri development, um, is

38:28 something that Kira was built in right from the get- go. Uh, and so basically lets you take a large complex task and break it down into like a plan mode before your agent actually gets to work on it. Um, and so I think depending on the complexity of your project, um, one recommendation to optimize on your credits is also to use spec mode first that your agent does does some of the planning before it actually does a bunch of building that you then have to call back and say, “Oh, I actually didn’t want you to build that. like now we have to delete all those files and we’re just spending credits to get work that we

39:02 don’t need. Um so I think specs is kind of interesting for some of those larger maybe more complex um projects or applications that you’re working on so you can help avoid wasted um implementation credits on approaches that you didn’t actually want agent to take. Um we talked about hooks. I think Eric, you mentioned hooks earlier, but uh I don’t use hooks a whole lot, but I know that I have peers that do uh a lot of hooks, which basically trigger your agent to do something while you’re in the middle of a workflow. Um, and I

39:34 wrote about here there’s some auto linting that you can do or like formatting for um your code writing depending on what it is you’re working on. Um, there are also hooks that I’ve seen for folks that um, rather than having like the agent write the summary markdown file um, in the middle of the project, if you are someone who is like, okay, once I actually check this repo into GitHub, then create a hook where now you can write the summary markdown file because we’re actually doing something with that code that we, you

40:05 know, want someone else to interact with. Um, so that could be something else that you can kind of work into your workflow is using hooks for some automations. Um, >> I’ll stop you there. I love I love the hooks idea. Just you can do so many creative interesting things with it. Like you mentioned having it kick off Husky >> rules or or lint rules or adding uh running test suites, pushing up PRs, creating commits, creating branches, like yeah, there’s a lot of interesting

40:37 things. So, anything you can do to like have it kick off deterministic workflows that don’t cost credits will save you time. We kind of had the similar uh last week we me and Laura were talking and our whole team was talking about Jev. You know, Jev is the big new hype machine everywhere. And what’s nice about Jev, it kind of reminds me of Hooks in some way. It’s Jev, by the way, if you haven’t heard of it, it’s this new type of model that gives you this probable pro probabilistic output. So you ask a question and it’ll give you

41:07 like they called nles. They’ll give you like a 50% or 100% or give you this percentage back of what it thinks it’s true and it costs very few uh tokens to use. But a lot of people are using that to then um have it kick off a deterministic workflow. Like someone created Jeb to uh had it ask questions and it would look up Wikipedia to find the answers. So, it had like a built-in MCP tool, but it would was able to take a question and then figure out

41:41 where in Wikipedia to look it up for. So, just give an idea just great great idea to to to look into to hooks and see what you can automate throughout it. Any kind of guard rails you can put in your code is a good idea just in general. And also, hello from Indonesia. Almond, thank you for joining. >> Hello Indonesia. Um, okay. And the final two here before I get to kind of showing this all, um, this one’s kind of I don’t know, it’s kind of funny and I think [clears throat] it’s sort of like common sense is sometimes not so common,

42:13 but yeah, I think for for a lot of folks like if you are able to do something simple in the code writing yourself, like if you are really strong with CSS, if you’re strong in JavaScript, like write write the actual code yourself. uh because if you can do it and really quickly just you know edit uh the weight of a font or you know something silly like that like you don’t need to rely on everything for the model. So I think that that was just like do the easy stuff yourself if you’re able to. Obviously that you know it it’s kind of in the name right like common sense isn’t so common but sometimes you know I

42:45 think we we often get so um reliant on some of the code generation that we forget that like a lot of the stuff is just a simple fix that we can do directly into the file into the editor um versus like relying on the model to do it all and then spend our own money to do that. So um and then the last one here which uh I think more people are getting comfortable with nowadays is really giving um the agents as much context as possible when you are trying to explain like how to do something. So I had a a demo a couple weeks ago where

43:19 we were building like a version of Flappy Bird that game where you kind of are moving like the little bird in between pipes and stuff. And one thing that we were thinking about was rather than having the model assume or you know create this idea for how the visual for the game should be, you can actually preede that by just sketching out something you know quickly on your desk or you know on a piece of paper or something or like taking a screenshot of uh a different game that you have in mind um where then you can just show

43:50 that context to uh to Kira in this case. So, either dragging in the file into the into the chat window or uploading the image um in your prompt. And I’ll show you guys what that looks like, but you can basically here in the little clipboard. Um this up a little bit. You can like attach a file directly to your prompt. So, whatever you are attaching uh will be added uh context for the model in that prompt. Um, so that’s something else that I mentioned in the article was like using uh I think hashtag file which brings up uh files

44:22 that you can attach and folders as well with hashtag folder. Um, so really just giving your your agent as much context as possible before letting it uh run tools that consume more tokens and consume more credits uh when you can just uh frontload that in inside the prompt. Um, Eric, anything you want to say before I jump into some of this? >> No, those are great things. I know there’s some tools, too, that you can use to run in your codebase

44:54 to like clean things up instead of using uh instead of like using your Gentic agent to do it instead. Uh, I was just looking some of those like uh you can look at you can run npexow it looks at your code. It’s like kind of a code intelligence open source project looks at your JavaScript and TypeScript files. It’ll look for like dead code ends. So some things you might in previously would have had to like prompted and use credits for Curo or whatever you’re

45:27 using to to to find all these issues, you can now just run these like static codebase intelligent tools to to save you tokens and do that um for you. And also there’s one called KIPP. It’s it’s another one that will kind of it’s specifically made for code bases that were written with AI tools and it just kind of finds some of those those issues. So, I wrote a blog post about it. That might be another way to save a few credits. But that no, those are all great tips. >> Yeah, that’s cool. I actually didn’t

45:57 know that that was uh something that existed out there. So, maybe I’ll have to take a look at some of those as well. >> Hello, Anton from Ukraine. Thanks for joining. We’re just talking about ways to save credits in Curo. Uh Morris written a blog post about it. >> Yeah. So the other thing I wanted to show here was I think a lot of this gets a little bit lost in the like abstraction of credits and prompts and like what does this actually look like in practice when you are doing it with Kuro. Um and I think the best way to

46:28 like really take a look at this is to run um to run your prompt with different models and really see you know what it is that the difference is. So, I have an example here of a project. I think I mentioned earlier um for those who are new, but let me move this out of the way. Um I was working on a video reel that was using an open source package that renders video from HTML that an agent writes. Um and so, uh yeah, basically creates like a a little teaser reel for a particular product or

47:01 a particular um I guess anything really that you you give it, it can create a little video for. Anyway, I was working on that and uh I wanted to take a look at like well what does it look like for Kuro here to actually generate a response to a prompt with different models. And so here I have the same prompt which is very simple but it’s just to write a two sentence summary of this project. It generated the response. The estimated credits used here were about 1.16 credits. And then the exact same prompt on a much smaller model uh

47:34 GLM5 is using one credit. Obviously this prompt is like not super complex. So you know the full power of the model is not being used here. But the idea is like same prompt. It’s costing me less to do uh a smaller model for the same basically result. Right? This is just a summary. Like I don’t need to be using uh opus 4.5 a frontier model to just generate some summary files. Um, and that’s the idea here is like play around with some of the smaller models. Take a look at what the different credit multipliers are and you know, think

48:06 about what you’re doing. If you’re doing a simple uh task, if you’re doing a simple prompt where maybe you’re just uh like we just mentioned, maybe cleaning up some files or doing something that doesn’t require uh like highle reasoning from a model, jump to a smaller model. Um and and and and take advantage of those smaller credit multipliers to, you know, extend the use of your full credit. Um especially when you’re on the free tier. Again, the whole point is like we’re talking about how to make those 50 credits on the cur free tier go further. Um, and so that’s kind of one way to to take a look at that. And then

48:40 employing all those different strategies that we just talked about, right? Like making sure that you have steering files. Um, let me put this down here. I’ll show you guys, but I have a couple of different um steering files that I have for different workspaces. Um, and that gets plugged into um each and every session that I have with Curo. So again, if I’m saying like don’t write um any tests unless you’re told to do so um files unless you’re told to do so, all that is going to save me in the end um when I’m working with the agents so that

49:12 I’m not generating those files that I don’t need to be. Um yeah, that’s kind of >> I remember the time I remember like early earlier versions like I don’t know if this was a model issue or what you would get so many markdown files. I don’t know, [snorts] anybody in the chat ever have that happen when you worked with a project and then all of a sudden uh like some markdown files started appearing. I feel like agents have done a better job of that. But yeah, anything you can put in that steering file that just says hey like yeah we don’t we don’t need to do tests on everything.

49:42 You don’t need to to do XYZ. You don’t need to write uh do minimum amount documents. Even with spec driven development when you’re using that there’s different spec driven modes now when you use it and this is in the ID and also it’s in V3 of the CLI too uh and you can be like hey don’t like just do a design document and don’t do the requirements document and just do the implementation so you can kind of play around with it as well if you don’t want it to go through all the the different steps. >> Yeah, >> I think quicksp spec is one of them.

50:14 >> Yeah, quick spec is one of them. uh which just kind of expedites the process. Um for those of you that are not familiar, when you do a spec file um with Curo, it will generate three files as it’s working through the plan with you. Um so it’s three markdown files, a requirements file, a design file, and then an implementation file, which is all the tasks that it believes the agent should um basically implement to actually get that project uh shipped. Um, so yeah, there are some use cases where if you already have uh like a

50:46 series of requirements or something, you don’t need to build a new requirements file off of that. Maybe just import a requirements file that you already have. Um, same for the other um two phases like don’t don’t uh rework the wheel basically is is the idea there. Um, one other thing I wanted to bring up, um, Eric, we talked about, um, this we have this open- source, um, benchmark, which is actually really interesting, um, that basically takes a look at like what

51:17 is it, uh, cost to actually run the same prompt across different uh, harnesses or or I guess agent AIS in this case. Um, and there’s a whole uh GitHub for this that uh was published and is being maintained. Looks like even two days ago was the last uh edit on this. But yeah, I would encourage you all to take a look at this as well. Um, the GitHub is called sample agent cost bench. Um, and the idea with this is like how much does it actually cost to use the same model

51:48 across different uh coding CLI or different coding providers or code identic coding providers. Um, and there are some early signals, uh, and there’s a whole video here that you can take a look at, but there’s some early signals that on Curo the comparison actually looks pretty good. Um, as in to say like it’s cheaper to run that same prompt in this case, like there’s a question here on SA 4.6. It was cheaper to run that particular prompt on Kuro versus Cloud Code versus Copilot versus Codeex. Um, and this is this uh, GitHub repo is is

52:22 like, you know, obviously free to use. is open source. So, I think it’s interesting, too, if you’re like wondering what it looks like to compare across different providers. Um, this this is something that our team has been taking a look at and um has been really interesting to to kind of understand. I don’t know if you have anything else to say on this. >> Yeah. Yeah, it’s No, that’s a good point. Like, we’ve uh I we also have a subreddit. It’s the Curo ID subreddit and we get occasional people coming in there asking about us comparing us to different models and the costs and we

52:56 definitely point people to this cost bench because I feel like we are very competitive with other providers. Ko comes out on top on a lot of them. So yeah, I would just check this this out. It’s a good comparison here and and but we we are definitely listening to feedback on on everything we hear especially cost we understand is top of mind for a lot of companies. The idea of token maxing there was this idea of like everybody is using as many tokens as they can. I think that kind of that’s kind of dead right now. You know everybody is thinking like how do we minimize our

53:30 token usage and not go crazy uh with it. So we that’s 100% top of mind for us uh everywhere with curo but I think we are very very competitive on this. Uh looks like I have a fire slop. Hello. Hello from India. Thank you LinkedIn user. We have a few more minutes left to answer some questions. Um one one question here. someone is hardcore cur user. One suggestion is to have a

54:00 different zish profile for cur so it won’t mix the with rappers use internal like oh my zish. Oh that’s a good tip. I I use a different um oh hi fire slop. Uh I I use what do I use? I I I was using bash for a while and then I use my zish or zsh for a while. Um and I haven’t had any problems. I set up aliases um to do different things to just quickly open up curo, but I haven’t Yeah, I mean definitely some rappers are interesting.

54:33 >> Yeah, one thing too I actually had an issue with this which I didn’t I I then found out was in the troubleshooting docs, but um let me see if it was in here. But yeah, I I use um oh, what’s it called? Um power level is it 1000 or something like that. uh as my as my wrapper. Um and I was noticing that it was it was breaking my curli. Um but yeah, there’s a somewhere in here in the troubleshooting is like you you have to uh kind of edit

55:04 the the zish profile so that to your point like it doesn’t kind of break uh with your wrappers, but you can still use it. Um it’s somewhere in here. I forget where it is, but >> is that that local shell integration thing? >> Oh, here we go. Uh there’s even actually one specifically for power level 10K users. So, yeah, it it it does in the troubleshooting docs for Curo, it’s a known issue um because obviously Kira pulls from your uh Zish profile and then it basically breaks the terminal if you

55:37 have it in the wrong line of the file. So yeah, if you are experiencing that, there is a section here in the troubleshooting docs for like how you should add that into your profile so that it doesn’t break the CLI. Um, this is the one that I have here for the Zish profile. So yeah, nice. Okay, says yes, aliases works, but let let’s say if my zish updates, then it’ll mess with ongoing workflows. Okay, cool. Thank you for using it. Yeah, good tip. Good tip.

56:09 Frag zero says, “Hello, I use cure every day and Kuro powers helped me a lot with MCB integrations.” Hey, love to hear it. Glad, glad to hear it. Um, awesome. Um, is there anything else, Laura, that you want to show us today? I think you’ve showed us so much. I think this is very interesting. Everybody definitely try out some of these strategies. >> Yeah. Yeah. No, I don’t have a whole lot. Uh there are some other like cool projects that I was working on this week. Um but I don’t think it’s like you

56:40 know super relevant to the >> Well, you can show us one of them. You showed us some earlier or >> Yeah, I mean I was working on is it this one? Let me double check. I think it’s my other window. Um I actually this week I was working on um we’ve been partnering with some Amazon devices um teams internally and they were asking me to kind of take a look at what it’s like to work on um using this uh new software for writing apps for um like Fire OS

57:11 TVs. Um so that was something that I was working on this week. But yeah, let me show you guys. Oops. Uh there basically is this like uh yeah this OS called Vega OS uh that if you are someone who’s developing applications for uh for TV um >> like Fire TV >> surfaces yeah Fire TV etc. um they they just kind of released this OS that helps you build um that kind of uh app and it comes with its own like virtual device

57:42 too. So you can kind of test it out. Um like an emulator and so like this is one example that I was working on this week with Kira is I had given it the GitHub repo for this um OS called Vega OS and I just was like kind of vioding with Kira to kind of get this app and running up and running. Uh but yeah, this was something that I was building this week. Um just to play around with like what it looked like to build some TV apps and oh my server might have gone down here, but uh yeah, this is like a TV interface that if it was working, it would kind of show you scrolling through, but I think I might have accidentally stopped the uh

58:14 the virtual device. Yeah, that was something I was working on this week. >> Yeah, it’s so fun. I was working on like this MP3 player. Um there’s so many neat things you can do with Curo. um just playing around with it. Cool. Well, I think we’re going to wrap it there. Thanks everybody in the chat that asked questions. Thanks for everybody joining and thank you Laura. I hopefully we can have you on more often and show more

58:44 things. I know you’re always thinking about Kirao and ways to optimize stuff and you always have some interesting topics to talk about. >> Yeah, thanks for having me and thanks for all the great questions everyone. Yes, thank you. Thank you. All right, we’re going to wrap it up here. Until next time, check out cure.dev and check out those workflows. Check out Curo ID and the workflows. Try it out. Take care. >> Bye. [music]