BONUS: Scott Hanselman Showcases Engineering with AI LIVE from Microsoft Build 2026

12 Jun 2026 · 44 min · 16 chapters

Ask about this episode

Ask anything about it. ChatGPT or Claude reads this page and answers with the times it was said.

Connect VO and ask about every podcast you hear, including the moments you saved. Add to ChatGPT · Add to Claude

In short

Scott Hanselman (Microsoft/GitHub) discusses AI-augmented software engineering at Microsoft Build 2026, arguing against “slop/vibes” by emphasizing SDLC rigor, speed limits, grounding in real data, and practical tooling on Windows/GitHub Copilot.

Guest backgrounds

Scott Hanselman is a Microsoft VP/member of technical staff focused on GitHub and developer tooling; long-time Windows engineer (since Windows 3.1). He builds AI-assisted workflows and demos “vibe-coded” apps evolving toward AI-augmented engineering.

Key claims

Use AI like a chainsaw but still do tests, code reviews, and read diffs. Don’t multitask too much (best 3–5 parallel worktrees). Avoid token leaderboards; use the right amount of tokens. AI must be grounded in data, not improv.

Notable examples

WSL containers/WSLC, Hansel Projects dashboard, “Hansel Projects” triage for Tiny Tool Town (GitHub-only database), Groot/TUI on Git, Copilot skills generating a grounded blood-sugar HTML report via Nightscout/Loop, and a grounded-vs-ungrounded chatbot demo (Family Feud vs real facts).

Written by AI. May contain mistakes. Listen to the episode to check what was said.

Chapters

Tap a time to open that second in VO

Scott Hanselman's Keynote Insights

0:45 to 2:10

Scott discusses his keynote presentation and the current AI landscape.

“Close them, they turn off, you open them, they turn on, there's no fan, everything's quiet.”

AI-Augmented Software Engineering

2:10 to 6:20

Exploration of AI's impact on software engineering and coding practices.

“So you can just type container images and have your Docker files work without having to install any third-party software.”

Managing Development with AI Tools

6:20 to 8:50

Discussion on using AI tools for efficiency in software development.

“Maybe it's this model Maybe it's that model like there's a slider bar of importance and I want to use just the right amount so here This is from yesterday I can generate this every day.”

Personal Projects and Emotional State in Coding

8:50 to 11:20

Scott introduces his personal project dashboard that prioritizes tasks based on emotional state.

“So if it's something like Windows, it might be a couple of weeks or a month because Windows has the Insider, they have the different rings.”

The Importance of Code Reviews

11:20 to 13:00

Importance of reviewing code in the age of AI-assisted programming.

“Like, there's room for what you want to do.”

Tiny Tool Town: A Showcase for Creative Tools

13:00 to 14:00

Scott shares his platform for showcasing fun tools created by developers.

“So we've got all these docs to talk about, well, here's how we want localization to work.”

Exploring Tiny Tool Town

14:00 to 18:19

Discover the concept and features of Tiny Tool Town, a platform for unique tools.

“So Tiny Tool Town is a place for stupid tools.”

The Evolution of Web Development

18:20 to 20:54

Discussing the impact of early web tools and the importance of creativity in coding.

“Like, I feel like this is where things move beyond vibes.”

Managing Content with JSON

20:55 to 22:35

Understanding the lightweight nature of content management using JSON files.

“I've got a little project like that that's that's essentially run off a JSON file connected to notion.”

Integrating Technology with Health

22:36 to 26:31

Exploring the intersection of technology and health through open-source tools for diabetes management.

“Yeah, so I run my blog, my podcast, Azure Friday, Tiny Tools, all these things in like an Azure basic one for like 45 bucks a month.”
Show all 16 chapters

Implementing Smart Solutions with AI

26:32 to 28:00

Using AI to enhance personal health management and improve daily tasks.

“And at its heart is a website called Night Scout.”

Introduction to Personal Voice Tech

28:00 to 30:08

Explore how personal voice technology operates without cloud streaming.

“Your personal voice never goes to the cloud.”

Using Skills for Blood Sugar Monitoring

30:08 to 34:25

Learn how to use skills to monitor blood sugar levels and analyze stress.

“Now this is a, I'm making this up, by the way.”

Grounding AI Outputs in Real Data

34:25 to 36:55

Understand the importance of grounding AI responses in actual data.

“You know, something I've been doing is, like, I'll tell it to visualize something.”

AI for Non-Engineers: Learning Journey

36:55 to 42:01

Discover how non-engineers can effectively learn and utilize AI technologies.

“but so much has been done with grounding in the last year, six months, to help even more accurately ensure that what it's using is right.”

Understanding AI and Coding Basics

42:01 to 43:44

Learn how to deepen your understanding of programming and AI tools.

“Think about where your level is and go, you know, I kind of understand HTTP, but I'm going to dig a little deeper.”
Hear the part that matters, and keep it.Open this episode in VO. Double tap your headphones to save a moment as you listen.
Get VO free

Transcript

Automatic transcript. May contain errors.

0:00To the Neuron AI Explained, I am Corey Noles here today from Microsoft Build 2026 in San Francisco. We're down at Fort Mason right off the water. I had a really cool keynote this morning and fresh off that keynote, we're going to be joined today by Scott Hanselman who works for Microsoft, works for GitHub. He's a VP, member of technical staff and he's going to join us today. We're going to talk about coding and Windows and all kinds of cool stuff as we go. Happy to be here. Whatever makes you happy. That's great. It's good to have you here. This is my actual computer that I work on every day, so we can do whatever we want to do.

0:33I asked if it was the Surface Ultra, and he said no. No, sir. This is my... But it is an ARM Surface. It's a Surface Laptop 7. It's my favorite machine right now. Big fan of the ARMs. I've really grown to like that. People have got to understand, these things are lovely. Close them, they turn off, you open them, they turn on, there's no fan, everything's quiet. It sits on my laptop for days, there's no wait. It's chill. It really is. It really is. So tell me, you all had a lot of big announcements this morning. And you came on and did a fantastic presentation toward the end. Can you tell us a little about what you went over?

1:03So we're in a really kind of crazy moment with AI right now. And I want to spend the next kind of thousand days as I get towards the end of my career really teaching people how to do AI augmented software engineering. And I feel like there's all this talk about slop. There's all this talk about vibes. But my degree is not in computer science. My degree is in software engineering. Software engineering matters. And now they've given us this power tool, this freaking chainsaw, and you know, and they're telling us that hey, it's just vibes, man. Just vibe it into production. But I think that AI augmented software engineering is more interesting than just vibing stuff.

1:39Like programming languages exist to be specific, and I like Windows. I've been working on Windows since May of 1990. Wow. When I installed Windows 3.1, and I learned how to do Hello World from a guy named Charles Petzold, who wrote a book called Programming Windows. And so, like, this is, gosh, what is that now? 35 years of Windows. 36 years of Windows, right? So I like Windows. Yeah. And I got to work on things like Windows Terminal, WSL, which lets you run Linux on Windows, WSLG, that lets you run graphical apps on Windows from Linux. Yeah. Like, there's cool stuff. We announced WSLC, which is containers.

2:15So you can just type container images and have your Docker files work without having to install any third-party software. That's amazing. So I like this. So I got to kind of move away from being a manager, and now I'm an individual contributor with that kind of generic title, member of technical staff. And in January, you know, I was playing with OpenClaw. I've been running my claw for a couple of months, and Peter was like, hey, you know, you made this. He saw my Vibe-coded Windows app. Yeah. And he's like, this is cool. You should keep doing that. Keep doing that. And I talked to the Windows team, and I was like, we need to do something cool here.

2:48Like, let's give me a couple of designers. Give me someone who's got actual taste. Yeah. And let's turn my vibe-coded app into something that is moving towards AI augmented. Kind of a middle of the road there. Instead of not just vibes, but not as rigid as software engineering traditional would be saying. Well, that's a great question. So like, what is rigidity? And I think it's less about rigidity, it's more about speed. So when you take something like being able to run in containers, being able to spin up an open cloud, being able to bring it back down, being able to put a PR up, run multiple PRs in parallel, have multiple copilots running in Git work trees in parallel.

3:24I think three to five is kind of optimal. Any more than that, you're fooling yourself. Going to start getting messy. People think they can multitask. No, you can't. No, we're not good at it. No matter how good we think we are. Humans don't know how to multitask. But I can juggle, though. Yeah. So I can do three or four work trees. So I use WorkTrunk. And I'll show you how all this works. And then I use the new GitHub copilot app. And then we can do a release a day if we wanted to. So we're learning as much as the folks working on OpenCLAR are learning. It's like, how fast is the right amount of fast?

3:53Yes. And how do you avoid being a slop cannon? Yeah. And you really be a software engineer, good SDLC, amazing tests, good code reviews, read all the code that the AI writes. Yeah. We are looking at the code. So I'll show you how I look at diffs. Oh, that'd be great. How we move through that kind of stuff. You know, it's really interesting because we very much are at a time where, not just the pace of AI, but the pace of development since AI is moving so insanely fast. And the ability to find ways to keep up with that speed in a competitive way, but not churn out hot garbage feels like the trick.

4:28So it's funny that you talk about the speed of stuff. Let me actually skip to something that you might be interested in that's kind of unrelated. So I mentioned how too many people are multitasking, and I think you see people token maxing. and like it's not i don't think it's cool to have a token leaderboard no no like you're lighting money on fire well you need to use the right amount of tokens to do the job so i'm not interested in number of prs number of commits lines of code yeah i'm interested in humans like what problem got solved what user story happened yeah but i also am noticing people that are just kind of like vibrating with enthusiasm and activity but did you actually do anything so i made this dashboard for myself because you can feel yourself getting that dopamine rush.

5:09Yeah. So this is called Hansel Projects and basically it looks through all of my different projects and it says pick the work that you have the energy for. So I like how am I feeling? What's my emotional state right now? Do I feel like doing PRs? Well I got 12 PRs. Well, but I got some issues on this one that might be interesting. But maybe... Am I feeling a little more problem-solving today? Yeah, exactly. Do I want to do human PRs? Here's some stale PRs I need to work on. But like this one is some like maybe some small automations. Yeah. So I don't want to like wake up in the morning, have a crush of a hundred to-dos, and then do all the bots and then go to bed and then rinse and repeat.

5:44So there's a humanity to it. So I made my own dashboard that looks at all of my repos. I've got 140 plus repos. These are different priorities. And it prioritizes it based on like my emotional state. Oh, I love that. And this is actually running in a... I'm doing lazy today. Yeah. Then that's okay. And we need to tell ourselves that that's okay. Yeah. Because like the social media kind of like leaderboard of people are like, what are you doing, bro? Why aren't you like burning tokens? Not interested in doing that. And you saw that in the keynote because I don't want to Burn tokens for tokens sake.

6:15I don't want to waste energy. I don't waste money Maybe some of the tokens are on a laptop. Maybe they're in copilot. Maybe it's this model Maybe it's that model like there's a slider bar of importance and I want to use just the right amount so here This is from yesterday I can generate this every day. These are open PRs actionable items, and you can see that I'm inside of the GitHub Copilot app running this local dashboard. This is the Copilot app. I used to spend most of my time at the terminal, so I'd have like here's a pane, here's a pane, here's a pane. So I've got like two PowerShells and an Ubuntu, right?

6:51And you can see over here I've got like all my different terminals from my OpenClaw, Ubuntu, I I could also install Kali Linux or Red Hat or whatever. Yeah. RSL is my main. But then I found myself spinning around console task switching. Yeah. But then, and I was running Copilot CLI. And now I'm inside of GitHub Copilot, the app, and I can see things from an actions perspective. What are the issues that are open? What are the PRs that are open? It's more stuff to be done than a terminal multiplexer. Yeah, I don't need a terminal multiplexer. I just need to know what's going on. So here I'm working on the open claw Windows node that we have launched today.

7:34I saw that. That's exciting. And you can see like I'm working on this thing called Mojibake and I had Christine back at the office fix that. Mojibake is when emoji gets corrupted because of double encoding and you end up having like an emoji or a Chinese character turn into like three or four ASCII characters. Yeah. So Christine has gone and fixed that for me and what's nice is I can see the changes that looks like a lot of changes but she was actually repairing a double encoding error right so there's there's the removal and there's the replacement back fixing the Chinese resources so I can look at this PR and manage this PR directly from here and then she's also gone and she's added 188 lines of test to make sure it doesn't ever happen again yeah so I'm basically spending a lot of time reading PR is looking at issues as a human and I'm basically applying my sense of smell.

8:26Like, code smell matters. People aren't reading this stuff enough, and I feel like that's important. I don't know, what do you think? I think you're exactly right. I think a lot of people just do things. And, you know, maybe when you're building a nerdy little app for home or something, maybe that's fine. But when you're dealing with real production software, obviously, that's a big difference. Like, how long does it take? So, she made these changes for you this morning. realistically when does that hit users? That's a great question. So it would depend. So if it's something like Windows, it might be a couple of weeks or a month because Windows has the Insider, they have the different rings.

9:03It depends because you don't want to go from, we'll use Christine as an example, from her computer to a billion computers overnight. So maybe a couple of months. And it goes through rings. Windows Insider gets it and then it finds its way. If it's my toy, like I wrote an app called Peak Desktop that brings the Mac feature where you click on your desktop and all the windows run out of the way and it lets you access your desktop and they come back. That's a Mac thing. I brought that to Windows. I can push that out and people will get it within an hour. I think there's a lot of stuff like that that'll really excite people that you guys rolled out this morning.

9:37Like all the new utilities. Oh yeah, that's a great point. So we're partnering with UUtils and they are Rust implementations of what people know is like the GNU utilities. And the idea is that if you're out at the command line you you want to fall into the pit of success. Hey, I just got this Windows machine. I've never used a Windows machine. LS. And you hit LS and you hit enter and it doesn't work. You're like, what the hell, man? Right? So we added like 75 utilities. So like grep should be grep and grep should work. And I shouldn't have to open, you know, Ubuntu just to get grep. Yeah. So that's the kind of stuff.

10:11Little paper cuts matter. Yeah. So we're fixing those and that's pretty sweet. And then things like the terminal, Like this is actually, like I believe, do you like Tuis? Are you a person of a certain age? Let me see if I can remind you of one of our favorite Tuis. And you tell me what you think. So a Tuie is a, here let's do this. We'll do a fun thing and you tell me if you remember this because you might be my age or you might be a crisp 29. I never, I can't tell anyone. 46. Okay, 46. So I'm pushing 53. That look familiar? Yeah, yeah. Reminds me of my old DOS days. Yeah, exactly. So this is a text user interface And TUIs, text user interfaces, are having a moment.

10:50And one of the things that makes TUIs possible, and fun things like that shadow and all these different characters, is a really rich terminal that understands all the things. That's the Windows terminal, which is itself open source. It's like this TUI right here is actually called Groot. And it is a user interface on top of Git. And you see, it's a new way of thinking about user interfaces. And some people are like, well, hang on, you're going backwards, man. No, you've got Windows apps, you've got TUIs, you've got browsers, you've got agents. Like, there's room for what you want to do. And we have an internet that's been built entirely on what humans like to build, how they psychologically interact with things, how humans best understand, best ingest information and respond to it as opposed to what works great.

11:41And the funny thing is it's kind of led to this sort of resurgence for the command problem. And I think that's okay. And what's great about it is that I can do my work here. So here I am, like, switching branches. I can come out to the GitHub Copilot app, and I can click on a pull request. Here's Christine's thing that she did a couple days ago. I can see the PR perspective. I didn't have to go to, you know, Copilot. Then Claw Sweeper, this is kind of cool. Claw Sweeper is a bot that's written by Peter and Vincent and the folks at OpenClaw. and it's going through here and it's saying, yeah, is this good?

12:13Is this a good idea? So it's done some of my review for me. And then if I want, of course, I could go up into GitHub itself, find those PRs, click on those PRs. See, PRs are already starting to roll in. And then I could even, let's see, here's a GitHub Actions one here. Here Regis did one where he's working on the gateway. I can look at this, and I could assign one of the other leads, or I could assign the reviewer. And then we're seeing kind of agentic conversations where he'll talk to Copilot and say, like, I disagree. I don't think that's it. Or Copilot consistently thinks our architecture isn't the way it should be, but we disagree.

12:50So we change the architecture documentation in order to steer it. Because AIs are really good at making features, but they're less good at making architectures. That makes sense. So we believe we know more. So we've got all these docs to talk about, well, here's how we want localization to work. So when Christine puts in her localization fix, we check it against that, you write tests. That's where you start moving a couple clicks from Vibes. Yeah. And into, like, we're actually making software for a reason here. I like that. Yeah. There's a formality to it. I think people, like, it's fun to, hey, man, it's just Vibes.

13:25Yeah. Is this for yourself? Great. Vibe. Knock yourself out. Yeah. Like, I made a stupid game called Baby Smash. We make a stupid one every time we get a new model to play with. You know, it's a silly game, right? So I go and I make this game, it's called Baby Smash, and it doesn't matter that it's silly. And it, right? So like, it's not a huge deal. But then the thing is, it got a couple hundred thousand users and now it's like, oh crap. Oh, now it's a thing. I made this for my kid and now it's a thing. So I ended up making a website that you might get a kick out of. It might be fun for your viewers.

13:58It's called Tiny Tool Town. Ooh. So Tiny Tool Town is a place for stupid tools. and it's made for an audience of one, which is me. But the idea is that all these people are making cool stuff. Why don't they get a homepage? Yeah. Right? So look at this. You can go and browse the tools and they're just piles and piles of fun tools. And here's the thing. Check this out. Let's look at here. And I think there was a coloring book that I saw. Hopefully Grant's throwing the links to these in the chat as we go. Oh, yeah. In the tiny tool town. So here you go. So Maria went and made a coloring book with the vibes.

14:32check it out it's called stillness you can go and pick your colors and like it's just something to chill out and color right that's just one example oh yeah some of these utilities are for the terminal some of them aren't and this is the part that I'm really proud of if you go to github and you say make an issue and you hit a new issue check this out submit a tiny tool it's a custom form. And you can submit your tool right there. Dude, there's no database. The database is GitHub. Oh my god. Right? Yeah. Okay, so this is like the unlock for me, because I vibe this whole thing, right? Yeah. So check it out.

15:13If you go over to the code and you look at GitHub and you go to workflows, so every day I get a text of the tools that have come in and then I look at them as as a human and I approve the ones that I want. So that runs approve tool, which is a GitHub workflow. And then from there, it goes and finds how many stars that tool has, whether or not that submitter is a person or a bot. It generates summaries. So for example, if we go look at her coloring book on the thing, this is all static, which is crazy. You can see it's all very, very fast. Come down here to the bottom, there's what she wrote. And then Copilot looks at it.

15:53and I'm very clear about what's AI generated and what's not. So then each tool gets its own kind of review. And even more, they can have their own separate pages. So every page looks different. See how this one has a totally different theme? Yeah. And then Tiny Tool Town itself has both light mode, dark mode, and GeoCities mode. GeoCities. Now it's a party. It should be stupid, right? Like, it's okay to just make the web weird again. How big of a role did tools like GeoCities and AngelFire play in creating the next 20 years of web development? What I'm trying to do is I want the young people and early in career people to feel like I think we felt.

16:34Yes. You know what I mean? Because I did this show called Azure Friday, and this website was made with Claude Opus. And let's just be real. It's mid. Right? It's fine. It's fine. Right? This website is fine. Like you can see the gradient under Friday and you can like yeah, that feels like Opus, right? And you know, oh, it's got light mode. Okay. There's things that it does that are cool like lightning fast search Mm-hmm, but until it has GeoCities mode with fire Right, it's just not the same man. No, it's not you got to look you got to sign my guest book and Join my web ring. Yeah, let's be fun.

17:15Let's do weird stuff And you know what? That was such a good way to learn back in the day, because there were so many things you were like, I could do this. And you needed to learn like HTML snippets, little things. Exactly. And before you know it, you're like, oh, I know how to do that on my own. Right. I know how to do that on my own. It was very, honestly, a lot like learning to code with AI is today. Instead of so much of a, let's sit down and look at a book and go through this very formally, it's almost like a, I need to do this thing. How do I do this thing? Exactly. So check this out. Let's go to Tiny Tool Town for a second here.

17:42Let's do a git. we'll do a get pull and and we will get pulled sorry and then we'll see what's going on so here we got some stuff and then I wrote a skill down here in the corner I want to zoom in on that for you and then I'm gonna say tiny tool town triage TTT triage and greater skills right right so these skills are gonna go off and they're gonna look at all of my stuff and in this actually I probably I shouldn't have used that model. So it's going to go and fetch the new tool queue. It's going to do all this stuff. Now this right now runs automatically, but we're going to do it manually. So there's one open tool submission that someone has done.

18:22It's going to look for duplicates. Was it done by a bot? Is it a good idea? Does it have a license? Yeah. Like, I feel like this is where things move beyond vibes. Yeah. Make a silly website for tools. That's cool. Yeah. But the thing about licensing, is it fun? Is it like people have tried to put commercial tools here? Yeah. I don't want that. So here we go. Public repo has images. Must be nerdy, yeah. Must be nerdy. Okay, and it seems to approve. It's local, it's MIT license. And I always, this is another good tip for folks who are doing vibe coding. Ask it it's confidence. Yes. But explain to you what your confidence is.

18:58I'm not gonna be confident if it doesn't have a picture, it's not done by a person. Did they make their account yesterday? Yeah, give it your own rubric of sorts. Give it your own rubric. And then I said, make me a table. And I'll just say, okay, cool, approve. and then it's going to go now and queue tool number 535 and then the best part is tool it's a balance let me switch away from the silly mode here and go back over to issues I've got agentic workflows that saved me time because this one that we just got in yesterday uh-huh the workflow checks for me does it have a license and it tags them because remember that labels you can get hub are totally flexible yeah so a human didn't have to go and tell me whether it has an image or not they can tell from the labels but that's toil yeah but i'm not going to make a thing auto approve it yeah human in the loop yes you got to do it you still got to have the human yeah man oh this is a good one i love the approach you took with like okay so i've got this thing i put it up it's a little bit fun holy smoke it's got 200 000 users i need to care now now and then like i think that's a really cool cool kind of use case in this whole let's prototype it and see if anyone cares Yep.

20:05And if they do, cool, I'll go back and invest some time. But if they don't, hey, I learned, I'm going to try. That's okay. And that's why I think one of these tools might take off and someone's going to do it and they found a home page for it. And if not, the whole thing is done on GitHub pages. Mm-hmm. It's done with static, what's it called? I keep forgetting the name of this. See, this is the other thing, that when you vibe stuff, you kind of forget. Astro, that's what it was. So I'm using Astro. It goes and it builds GitHub pages and then all the images get pulled in automatically I'm like everything I could do to make it as lightweight as possible You don't always need a database.

20:41Yeah another thing I would tell people you don't need to super base for everything You don't need to go and spin up a cosmos DB. Yeah, this whole website is run entirely off of a giant JSON file That's wild, but it's but is it though? Like how big would a JSON file need to be before you think it was too big that makes sense I've got a little project like that that's that's essentially run off a JSON file connected to notion. Yeah. You know, it's like, I mean, it's kind of a database, but it's not kind of. So what is a database? Right. So a giant spreadsheet. So look, yeah, exactly. Right. Most of the world runs on Microsoft Excel, right?

21:16So if we look at my blog, which has been around for 24 years, right? So I've been blogging since 2022. What do you think the database is for this blog? First off, wow. Way to go. 2002. Yeah, man. Yeah. I'm impressed. People of a certain age. So what do you think the back end is? Like we got 22 years ago could be access, SQL server. Could be Excel. Yeah. Could be 5300 XML files in a folder. It sure could. Yeah. So it is. There's one each blog post each day technically is an XML file. Okay. Which sounds insane but XML was the hotness before curly braces became the hotness. Yeah. Doesn't matter. By the way folks I'm not sure.

21:58It doesn't matter. It can be spaces, it can be tabs, it doesn't matter. It can be curly braces, it can be angle brackets. It's not that deep. Yeah. Structured data. It's... You're overthinking it. Everything is name-value pairs. So I load, when this thing starts up, 5 ,000 files up into memory and I cache 22 years of my writing. None of it was AI generated. How many megabytes of memory do you think that the 22 years of my life's work takes up as I cache everything to make my website lightning fast? I'll bet it's an embarrassingly small number that makes you feel like you've done a lot less work than you think you have.

22:33Yeah, it's 300 megs. Oh my gosh. It's 300 megs. 300 megs. Yeah, so I run my blog, my podcast, Azure Friday, Tiny Tools, all these things in like an Azure basic one for like 45 bucks a month. You could store that on... You can do it on a Raspberry Pi. Oh yeah. Right? So like look, here's a blog post from 2009 where I I was at Queeznos and I made an app where you could tweet your sandwich order. Because we were trying to like, this is before, gosh, before all the things that Uber Eats and the Instacarts and stuff. So we made Tweet Sandwich. And this is in 2009. But if I go and I pick another blog post, it's in memory.

23:14It doesn't need to not be in memory. No. This is all Jason. So I've got a podcast called Hansel Minutes because I'm not a very good estimator. and people would always ask me how many minutes that's gonna take and I would tell them and they'd be like, those real minutes? Those Hansel minutes? That's fair. So the show has had 1 ,046 episodes. Wow. So every Thursday for the last 20 something years. Giant JSON file. Wow. You pick a picture, you say open image and new tab. It's on a CDN, episode number dot JPEG. Like, you're overthinking your architectures. Your back end does not have to be that complicated because the show comes out every Thursday.

23:57Yeah. I don't need a whole database. No. I know what it is. It can be listed anywhere. Yeah, yeah, yeah. But then I make a database. Do I care about the database? I made, let me see if I actually have my password. I probably won't be able to do it on a live stream. Good catch. Good catch. I vibed a front end for it because it was just a JSON file. Yeah. So I made a cool little Opus 4.6 front end, and it makes it a little nicer. Nice. But what I ended up doing is less and less do I log in, and more I just tell my agent to do it. Yes. Because then you make a nice API. So all those REST APIs that we took the time to make over the last 10 years make a skill.

24:35Actually, let me show you something. That's what I do too with the site I was talking about is I have it just, you know, go make sure all of this is current. And come back, give me a report of anything that's questionable, and then we'll go through it, and I'll be like, fine, go update. Yeah, yeah, yeah, yeah. Actually, let me see if I can show you something that might not even be on this machine, and that might be cool. So we could actually do something live here. So let's go and look at my Copilot Skills folder, and we can see I've got my expense reports and my code review and stuff like that.

Read the full transcript

25:03And let's try. I'm going to go to the temp folder here. This will be fun. Trust me. And see if there's anything in there. Okay, no temp folder. Perfect. So I'm on D. I'm going to go ahead and zoom in for folks. We'll go to GitHub and we're going to go to temp and I'm just going to be in an empty folder here. So we're going to go and make their temp. And then in this empty folder I'm going to go and say copilot and I'm going to say dash dash YOLO because you only live once. That's right. You only live once. We're going to bring copilot to dash dash YOLO and that's going to bring up the copilot to your life.

25:34Now this machine hasn't been on for a day so it came up. I thought it was going to update. And I'm going to take the model and I think we can probably take it down. We don't need to use the top of the line. We don't want to waste tokens. Okay, cool. And then I have over on my website a thing called Night Scout. Now, for context, I'm a type 1 diabetic. I've got an insulin pump in my arm right here. Okay. And then I've got a sensor in this other arm. It looks a little messy, but it's stickers. So this here is the read. This is the right. So this puts insulin in me, and this tells me what my blood sugar is.

26:12And traditionally I have to take that information and I have to read it, look at a Bluetooth thing, make a decision, and then tell it to inject in my arm. So they don't communicate with one another? They don't communicate with each other. What do you think there's a strategy there? So the strategy is open source, right? Yeah. So there's a product called Loop. It's open source and part of the open source artificial pancreases community of what it's called OpenAPS. And at its heart is a website called Night Scout. And Night Scout is this cool application that lets you take the data from your, this is called a Dexcom.

26:46It goes from here to my phone, from my phone to the cloud and into Mongo. Ooh. Okay, and once it's in Mongo, you're like, okay, what can we do? It's gonna be crazy. Yeah. It's gonna be something wacky we can do. So if we go out to the command line, like out here, you can see the two most important things in my life, the current git branch and my blood sugar. Wow. That is my blood sugar right now. Wow. Okay. That's the version of.NET. That's how many work trees I have. That is the current folder. And there's my blood sugar. So I've got a REST API. So now I have a REST API. Hang on. Maybe I could make a skill.

27:24So here's a skill. Let's take this skill. I'm just going to put that in my clipboard. And let's drop back out to the command line, and go back out to Copilot. And then, now my hands don't work as well as they used to. I've had all kinds of hand surgeries and stuff. So I'm going to use dictation. Okay. I want you to take this GitHub repository. It's got a skill. Bring the skill down, install it, ask me what you need to ask me, and then we're going to check my blood sugar. Okay. Now, I could, I was using Handy. I could use Windows H, which is built into Windows. Yeah. Or we have a thing called Slash Voice that's built into GitHub Copilot.

27:58Okay. And we actually use a thing called Nemetron. happens locally, doesn't stream. Your personal voice never goes to the cloud. Parakeet? I was using Parakeet just then. Yeah. Exactly right. Okay. So then I'll paste that in and I'll hit Enter. And it's going to go and take that and I'm going to expect it to bring that skill, fetch the repository, it's going to see how the skill is packaged. The skills folder is right here. Because I've got Windows Power Toys, I'm going to put that always on top. And we'll just see if - You need some haptic feedback in your right arm. Yeah, I kind of do, man. I need something to tell me when things are going on.

28:33Yeah, this thing will click, actually. Okay, so now it's asking for my website. So I'm going to use a website called Type 1T. I'm Type 1 diabetic, so Type 1 Pal. And Type 1 Pal is going to have me log in, and I'm just going to grab my URL, the URL that I'm going to use for my night scout. Now, if I remember correctly, actually, See how it's asking me a question here? Yeah. Co-pilot is requesting information. I'm going to go out, and I think it's in my PowerShell. I'm going to use Microsoft Edit and look at my profile. Microsoft Edit's new. Yeah. And there's just my... Okay, so it looks like my NightScout URL is in this environment variable.

29:18Okay. So I'm going to take that, and I'm going to say... And see how I'm clicking in the TUI, so you have mouse support in the TUI. I'm going to take that and I'm going to pipe it. That's important to you. That's awesome. Yeah, that's kind of cool. So I'm going to put that in the clipboard. So I just took that URL and I put it in the clipboard. And now I'm going to go back over here and we'll paste that in. And then we'll say there's no secret and the glucose units are a certain way. So now it's going to go off and it's going to call that JSON API. It's going to install the skill in my local thing.

29:46It's going to read all these different things. It's writing a little bit of Python. It's writing a little bit of PowerShell. I'm gonna expect now to see Night Scout. CGM means continuous glucose meter. That's what this guy, the read. So it's reading constantly. Every five minutes. Okay. It goes and it Bluetooths, hey, there's your blood sugar, hey, every five minutes. Yeah. All the time. That's great. I keep my phone on all the time. Okay, cool. My current blood sugar is 94. So that's cool. So now, let's look at this. Now this is a, I'm making this up, by the way. That's okay. That's okay. That's okay, yeah.

30:21Make it up. I'm wondering, I was on stage this morning, and it was about what, about 10.30? Yeah, roughly. 10.30, maybe 11. Okay, so I was on stage doing a presentation this morning at about 10.30, 11-ish. Make me a chart or some kind of a table, and let's see if I was stressed out on stage. So let's use that skill and give me that as a markdown file or something. Or a markdown. Maybe make, actually no. Dude, sorry. Non-gender specific dude, check this out. Make me like a like a cool local website report using that and and use the night scout skill to do it That'd be hot. That'd be hot. Yeah, I'd love it So you sound a bit of a yapper.

31:00Yeah, it's okay all that context kind of doesn't hurt Yeah, because you're just like chubby my sugar. That's not interesting. Yeah, give it to me to try. No, no, no Yeah, it's okay. Let's do it this way. Let's do that. Oh, here's a better idea exactly. The ADHD just rolls through. Yeah, ADHD So here it's using some of the built-in Python scripts that I've got and it's asking it, well, I guess he said today. Now while it's doing that, if we go off and open up that CGM skill and open up scripts. And I think a lot of people don't realize every time you're seeing one of those, that's a PowerShell window.

31:34Like it's just operating in PowerShell repeatedly. Yeah, that's a great point. This is a great point for folks to remember is that if you're on Linux or on a Mac, It's going to maybe use bash or ZSH. In this case, it uses PowerShell. All kinds of things that it could potentially use. It's very command line heavy. It's a word I should have used. And that's a very good point, though. It doesn't have to be command line heavy. Because, for example, if I want to jump in and look at the actual Python, this is the Python, I could look at it. If you find the command line intimidating, that's totally fine.

32:05If I wanted to do it from the GitHub app, I could go and open up that folder and say start that session from a local folder, and I can see it graphically. If I wanted to click over here and look at those files inside of open folder, we were in dgithub temp. Boom. There we are. Data's coming down. Here's some blood sugar. These are the JSON files. That's great. Right? So let's go and see what it's coming up with. I'm writing a local HTML report as a single self-contained file. Single self-contained file. Right. And this is the other thing I think that people don't understand, and I think this is cool.

32:44There's vibes, but I think there's a new kind of just-in-time UI generation. Yeah. Like, this is taking long, but if I was maybe using one of the new powerful computers, if this was not this machine and using the cloud, if it was using, like, a Surface Laptop Ultra, what if I just said, hey, what's my blood sugar? And let's say that this is, like, a TV, and I just say, put it on the TV back there. Yeah. And it's somehow connected in, like, a canvas. Yeah. Throw it up there. In a Project Solora kind of way, maybe. Well, or like Star Trek. Yes. Ultimately, all of this stuff is Star Trek. Yes. You remember, well, this might date me.

33:19So in 1984, I went to the opening of Star Trek IV. Nice. I met Uhura in Portland, Oregon. Ooh. My dad pulled me out of school. Good dad. And it was fantastic. Yeah. And then there was a scene where Scotty is talking to a guy says, come and tell me about this thing. He says, come to the Max. And he goes, uh, computer. And it's like a little Mac. And the guy says, just use the mouse. And he goes, oh, a mouse. How quaint. Picks the mouse up and he goes, computer? Like, they've been promising us that from the 80s. Look at this. They just did this. Was the stage stressful? Like, we just thought, I didn't come here to hang out with you and have this planned.

34:01That's right. We just made this up, like, just now. Verdict, maybe a stage adrenaline bump. Okay, so it looks like my blood sugar was a little high. It went up 35 points during my time on the stage and then fell back by dinner. And it made a chart? Yeah. Bro, you know what I mean? If I had given it more context, it could have put a lobster, like, at the peak. Yeah. And it said, it's Peter's fault. You know, something I've been doing is, like, I'll tell it to visualize something. 100%. And be like, you know, visualize a transformer. Yeah. Explain this. and it'll build up this whole HTML page that it'll bring up and you can see things moving and interacting.

34:41And I often wonder if some of this isn't because it's not necessarily using what would have been traditionally right, but it's maybe using what is the most efficient way to do this. Yep, exactly. You know, that meets the goals and is as efficient as we can do it. And sometimes that's an answer that a human might not have chosen. Yep. Yeah, and another thing that's worth pointing out is that these tools can be reused so once it's done a thing you can say oh that was good yeah save that yeah and this is important because I don't want to run an LLM I don't want to waste energy I don't want to you know spend money unnecessarily if this doesn't need to be an LLM anymore I need to move it immediately into a Python script yeah so generating a report is useful enough that I've gone and wrote Python to do it yeah so now I've got time and range distribution of my blood sugar.

35:32Look at this. Oh, yeah. That's not vibes. That's me saying, I need a heat map and saying, and then, oh, and I want one for my doctor and generate me like PDFs or whatever and make it so I can go and print this. And now I can print it out and hand it to my doctor. I should also add that in the last 10 minutes, you've just created something that is a level of healthcare monitoring that individuals don't have. Yeah, that's a great point. And they should yes and I think that they should be available to everyone yeah and then for people who might be concerned that like is this hallucinating these numbers I can prove that it's not because I have access to the data and I can go and see immediately using those if something feels off or you're like that's question but it's grounded yeah it's grounded in the data and this is important let's look at how it actually grabbed it the data was grabbed literally from this Jason right here, which I can see, which came from that API that we called.

36:27Yeah. And, yep, I can see the URL. I can click on the URL. And that's what was connected directly to your arm. And that's connected to my arm. And look, serum glucose value is 110. That is reality. So it's grounded in this database, and right now I can prove that and go out to the command line and hit enter, and it says 110. You know, I think that's a thing when it comes to AI that a lot of people don't understand yet is we still talk a lot about hallucinations, but so much has been done with grounding in the last year, six months, to help even more accurately ensure that what it's using is right.

37:07Yeah. And that you have an answer you can trust. And there's been a lot of that that's come in from like the enterprise software world even, where they've been trying to figure out how can we make this business friendly? Yeah. Actually, as we are running out of time, let me give you one quick little local thing that you might get a kick out of that we can talk about. Let's see this. Let's try this. See if I have node in this machine. Otherwise, is it npm run dev? And then server. Is it node run or run server? I always forget. npm run server? Okay. Shoot. Remind me how to do that. So I've started that one.

37:41That one is... I want to show you something real quick. Let me pull this off. npm run dev. That's one server running on local host. I need another one running over here. There it is. So I'm on port 3001 port 3000. So I'm going to hit this. I'm running a local chatbot. Okay. Test, see if it works. It's live TV, right? Okay, cool. Complete this sentence with one word. It's a beautiful day. Let's go to the. Okay, so what's the right answer? park Well that that's because you're a human you have common sense. Yeah, I actually built a probabilities heat map here So it turned green. Well, it could have been beach.

38:23There was a two point. Yeah, six percent chance There's a parallel universe where you know we're going to the beach now. Oh, what's this? I'm gonna hit it again Beach why because now parks in the context. Yeah, right. So is it grounded in data? This is not data This is this is family feud. Yeah It's exactly Family Feud. So watch this. Watch me tell it. Watch me get exactly what I want. Complete this sentence with one word. My name is Scott Hanselman. I'm from Portland, Oregon. My dad is a firefighter. My mom is a zookeeper. It's a beautiful day. Let's go to the... Zoo. Right down the primrose path.

39:01There was a 1 % chance that it was going to go... Mountain. To mountain. You know what I mean? And mountain's a great idea, too. But look at that. So when you do improv, it's yes and. Yeah. You know? Yeah. It really wants to do yes and. And the way that you can lead it should remind everybody who's watching that these things are improv partners. They are not fact machines. And that's important to remember. But when you ground it in real data, when's Brad Pitt's birthday? Okay, that's a fact. What is today?

39:39Yep. It's funny. Okay. Watch this. It's even worse. What's Scott Hanselman's birthday?

39:47None of that is right. It's not even close. Now watch this. That's not even close. Try again and do better. It's going to just keep making updates because Brad Pitt is famous, and I'm a random YouTuber. You see what I'm saying? You've got to ground that stuff. And rather than saying, I don't know, It's going to improv. Yeah, and that's why grounding in enterprise data and grounding in real data and grounding in my blood sugar in a Python script that I wrote matters and for an individual I think personal memory is going to play that same role 100 % like as it learns more about you and and and gets feedback from what you're doing over time You know the more that happens the more accurate it's going to get when you ask a question, right?

40:30It's also going to be less accurate if someone else comes and uses your AI It's probably going to be garbage for them look at that But it had all kinds of thoughts about when my birthday is, and none of them are my birthday. None of them are my birthday. Not even the month is right. Wow. Yeah. Well, one day I'll be as friends. It's like, well, the internet doesn't know. We'll just spit one out. I don't know that guy. Yeah. Yeah. So those things are important for us to know. So I think that as we are excited about Vibe Coding, we need to remember that AI augmented software engineering involves driving stick shift and understanding how computers work.

40:59And it's just a wonderful time to Vibe, but use them to teach yourself. Use them to get your brain to expand. Before we go, I'd like to ask one question. Yes, sir. How would you, say you're talking to someone like myself, who does not have a traditional software engineering background and has very much come from my degrees of philosophy and journalism, and everything I'm doing with this is stuff that I have learned. In using AI in a way to teach me, what would you say is the right approach for someone who is new to this? Like, what should they really go learn still? And what should they work their way through, do you think?

41:33The basics. Like, I don't know, because I'm a person of a certain age, when I got the opportunity to get a driver's license, my dad said, you can't drive this car until you change your oil, you change your tire, and you learn to drive stick shift. I'm that age. But if you force people to do that, then a lot of people would never drive. Yes. Because you'd be putting, so if I said, well, I'd love you to learn assembler. Yeah. And then learn C, and then go to, that's making creativity have a barrier. Yeah. So I understand that, like, Uber has value. But think about what you know. Think about where your level is and go, you know, I kind of understand HTTP, but I'm going to dig a little deeper.

42:10So take your comfort level, and I would encourage you to pop the hood and see if you can change the oil. Learn HTTP. Learn DNS. Jason. Jason. Learn the basics. Learn the drive stick shift. And you don't have to drive stick shift every day. Yeah. But it changes your relationship with the vehicle. That's kind of what my learning journey has been like. It's very much like, I have an idea, and I want to make something that's fun. And I'll go into, like, let's say, Codex or whatever. and I'll vibe something out. And then I want to understand how it works, why it works, and I'll take that code and throw it into another AI and be like, walk me through this.

42:40What do I see? How is this working? And help me understand things and have a back and forth conversation with it because you don't want to put something out if you don't know how it works. And the truth is, in doing that, there have been a lot of things where it was like, explain what, like a lot of people don't know what GitHub is. Somebody coming to this from nothing is like, what's a GitHub? Yep. You know, and like that's a conversation in itself to be had. That's a good thing to learn, by the way, if you're starting. So when I make docs for my apps, I do things like junior engineer specific docs.

43:09Draw me a diagram. You said you have a background in journalism? Yeah. Interview it. Yeah. Talk to me about what you were thinking when you made this code. That's great. And so I'll make the thing, I'll go for a walk, and I'll put it in voice mode, and I'll switch, I can also switch, by the way, GitHub, and I can switch it into remote and say, well, I'm gonna enable remote control, put it on my phone, go for a walk, and and steer my machine while I'm off in the sunshine. Wow. That's really awesome. Yeah, that's a good time to be a developer. But I think people just need to like, hey man, the vibes are immaculate, but you'll be better if you really become an AI engineer.

43:44Yeah, I think you're right. I think you're spot on. Scott, thank you so much for joining us today, man. This has been an absolute blast. I'm glad. Thanks for having me, man. To everyone watching, please take just a moment to like and subscribe to the channel if you haven't yet, and go through and be sure to check out the other videos, check out the website, and check into all of the awesome releases that Microsoft had today. You can't wait to hear about Project Solara. That looks really, really awesome. Same with Surface Ultra Pro, the seven new AI models. So many things. And I think it's really worth taking a look.

44:17They came to play this year. Thanks for watching. See you next time, humans.

From the publisher

Live from Microsoft Build, Corey Noles sits down with Scott Hanselman for a hands-on Neuron LIVE episode about AI-augmented software development, how it differs from just "vibe coding", and the surprisingly practical things people can now build with tools like GitHub Copilot and more.


Scott is one of the best technical explainers in software: a longtime Microsoft and GitHub developer, teacher, speaker, author, blogger, and podcaster who has helped millions of developers understand new technology without making it feel impossible to learn.


This episode turned into a live demo tour of what AI coding can already do, led by Scott's own use-cases. Corey and Scott walked through a series of examples showing how AI can help people build useful apps, prototypes, workflows, and small tools from everyday ideas, including Scott's own vibe-coded tools Baby Smash (https://www.babysmash.com/), which lets babies press random buttons for fun shapes and sounds, and Tiny Tool Town (https://www.tinytooltown.com/), which showcases random, cool tools Scott found around the web.


But in the coolest demo of all, Scott shows how to take an open source tool and create software a personal blood sugar tracking app for his own diabetes management. If that doesn't get your idea blood flowing for what you can do with AI, we don't know what will!


https://www.theneuron.ai/

More from The Neuron: AI Explained

All 106 episodes
BONUS: Scott Hanselman Showcases Engineering with AI LIVE from Microsoft Build 2026The Neuron: AI Explained · 44 min
Listen in VO