953: Why V0 Creator Left Vercel to Fix GitHub  ## Metadata - Author: [[Syntax - Tasty Web Development Treats]] - Full Title: 953: Why V0 Creator Left Vercel to Fix GitHub - Category: #podcasts - URL: https://share.snipd.com/episode/d16d2a2a-296f-4736-9603-4a54b4f57ca9 ## Document Notes ## Highlights - **Continuity Is The Future Of Developer Workflows** - Developer workflows need continuity across tools to preserve flow state and avoid context switching. - Jared emphasizes tight handoffs between VS Code, GitHub, and Azure to keep tasks seamless. Transcript: Jared Palmer Think that we need to do a better job as a community, building out continuity, if you will, right? Yeah. It still feels like, remember before, it still feels like maybe, I'm going to date myself, in high school when you worked on a word, we didn't really have like Dropbox, right? And like the pre-dropbox days. I know that it's like dating myself there, but it does kind of feel like that where these are so isolated. They're not fluid. There's no continuity. Your flow state gets broken. You start one task and you go off, you kind of forget yeah it's called final.js drag and drop underscore underscore final final yeah yeah totally so so anyway so i think that in the future And ideally one of the things i'm focused on is bringing continuity across all of the the microsoft if you will touch points, which is like, and that's what's been so important about This reorganization around the core AI unit, which is Visual Studio, VS Code, GitHub, and Azure AI Foundry. Okay. And parts of Azure too. And the idea here is that like, we should be working really tightly with the VS Code team at GitHub. And we should be making handoff super seamless. Right. Right, you know. And so when you start tasks in one, you should be able to pick them up on the other and vice versa. And you're seeing that today. And the other thing is that model choice is real. And I think developers want that choice because they always want to use the bleeding edge. And so that's one of the announcements we had today that we're bringing third-party models to GitHub Copilot and GitHub Agent HQ, which is the new control plane from which you can kick Off kind of background tasks and work through PRs and other things like that. ([Time 0:01:50](https://share.snipd.com/snip/a4d96873-5976-4c4f-bcaf-ee06755a3182)) - **Open Agent Tasks Directly In VS Code** - Open long-running agent tasks in VS Code to maintain your flow and edit PRs interactively. - Jared used Agent HQ to upgrade a Next.js site and then opened the draft PR in VS Code to finalize changes. Transcript: Jared Palmer If you do start a task in Agent HQ or on GitHub, you can open that in VS Code. Wes Bos Yeah, that was a killer thing to me. I think that was the most encouraging new feature shown today, just because like we were saying, it's like so many times when those responses come back, if it's unusable at that point, You're like, okay, now I'm 14 minutes. Scott Tolinski You're stuck. Yeah, let me scrap this thing. Or I scrap it entirely. Let me type in the box again and let it go run for 15 minutes and then come back to it. You lose that flow state. Jared Palmer Totally. One of the things that I did testing that feature, so last week, Next.js announced version 16. Very exciting. You guys probably covered it. If you didn't, whatever. But I was excited. Scott's a huge Next fan. Yeah. So the first thing I did was, and I thought this was pretty cool because one of the cool things about AGN HQ is the prompt is actually set up to create PRs, which is a little different than Like the VS Code, super interactive, fast Twitch, if you will, like cursor. Yeah. Not background agent, but like cursor sidebar situation. Yeah. The prompt that I gave was research the latest Next.js 16 release and upgrade guide, then run all the code mods and upgrade my site. Mm-hmm. And it just did it. And it created five different commits. It created a draft PR. And it did take, I think it took, I don't know, 10 minutes. But that's okay. But it did it. And then it ran all the, it not only did that, it ran all five code mods, migrated from NextLint to the new ESLint CLI. And it verified, it migrated to TurboPack, changed configuration, validated the builds, even included like now built in nine seconds of TurboPack. ([Time 0:03:36](https://share.snipd.com/snip/6298417a-f695-4cd9-9ad3-0fd1e68254ac)) - **Give Developers Model Choice** - Allow third-party models in Copilot so developers can choose models that fit their needs. - Jared argues model choice improves product quality and aligns with GitHub's mission to empower developers. Transcript: Jared Palmer Yeah, I think it's important that we, from my short time at the company. Yeah, I was going to say, 13 days. 13 days. But obviously, I talked to Satya before joining, and we had talked about some of these big decisions, right, for the platform, whether we should, like, allow third parties and stuff Like that. And we both agreed. I think it seems like this past, at least, like, three to six months have shown that, like, giving model choice is what developers want. It is ultimately leads to the best product. And, you know, if you're talking about GitHub's mission, which is to empower developers to build awesome stuff, well then like, yeah, they should have model choice. It's kind of a no brainer. ([Time 0:06:18](https://share.snipd.com/snip/5a259cd0-bea9-40c4-8302-254e09fbdccc)) - **Pull Requests Will Change With AI Coworkers** - Pull requests must evolve to reflect agents acting as coworkers, not just autocomplete tools. - Jared expects GitHub to redefine what a pull request looks like by 2027. Transcript: Jared Palmer What is a pull request in 2027? And I don't know the answer yet, but I'm pretty sure GitHub is going to have to answer that question. Scott Tolinski Yeah, yeah, yeah. It might look a little bit different. Jared Palmer Might look a little bit different. And that's exciting. And I think that's where we're driving and focusing, if that makes sense. That's cool. And then also to the performance side, this is something that's near and dear to my heart. So we're also working on that, too. That's been in the works. But I'm really excited about the team and the stuff that I've seen so far. Yeah, but now you get to show up and take credit, right? Because that's your thing. Right. Show up and take credit. Yeah, exactly. I contributed to not as much as you could expect for this, but I'm very happy and proud of the team. They put in a lot of effort. A lot of the stuff you're seeing today was started months and months ago, obviously. ([Time 0:08:51](https://share.snipd.com/snip/a5aaef6d-53a3-446b-8f90-6c1bd7b7a32c)) - **Scaffold First, Then Research With Agents** - Start projects by scaffolding (e.g., Next.js) and preinstalling tooling like Biome and ShadCN components. - Then prompt agents to research the codebase and write a graded plan before executing changes. Transcript: Jared Palmer To answer the podcast question i think i would do the same thing i think i would start out with like my I would scaffold an XJS app, biased, slightly biased. I would, but you know, SSR, it's important. I would get like biome set up and like everything, like to that level, if that makes sense. I would install ShadCN, I probably would pre-install the subset of the components that I thought would be useful in the UI. I actually did this, by the way. This is a specific question. For the old undefined site, I actually use that sometimes. It does like a sort of demo thing. And I actually migrated. It hadn't touched the site in so long, so it was in Gatsby. And I did the migration to Next. And what I did was, I think the Tailwind guys did a podcast template in one of their... Oh, yeah. Which I bought. But then I copied that thing into a folder. Tailwind UI, yeah. Yeah, yeah. I copied that thing into a folder. I downloaded a folder. And then I was like, yo, look at this folder. But one thing I'll give as a tip for those who are prompting, it doesn't matter what, which one of these agents you use, a trick that, I actually picked this up from the Cloud Code team that Works really, really well is to prompt, even before you either write a spec or actually go to execute, is to ask the model to do some research. Wes Bos Yes, yeah. It's so powerful. It's been powerful for me. And even like getting into like database design or those types of things to pre-research. Jared Palmer Like go research this. Yeah. Like get a sense of my code. Even that, those two prompts, go research this and get a sense of the code base or paste in a bunch of links and get context. Packing that context can be so powerful. And then either, you know, write a plan. ([Time 0:11:40](https://share.snipd.com/snip/c15af5fb-8982-4674-9da8-613b8a4a21f4)) - **Prompt Tricks To Improve Agent Outputs** - Use prompt techniques like “ultra think”, add logs/asserts, and ask models to grade specs with a rubric. - These tricks improve reasoning, debuggability, and produce iterative, reviewable plans. Transcript: Jared Palmer I don't know if you guys know this. No. Okay, so have you heard of ultra think? Scott Tolinski Yes. Jared Palmer Okay, so you can say ultra think and that forces usually force, at least in the anthropic models, will force like high reasoning. Okay. Another sort of secret tip I'll share, which I found useful, something like emergent behavior, is if you ask these models to add logs and assert statements, which is like low key, like Not many people use these days, but add assert statements and logs to maximize debuggability. Yeah. That can sometimes be very helpful when you go to like hit it out. And the third one that I found interesting, and this has helped me with planning. When I'm getting it to write a spec, I'll put that in a markdown file, like to do or plan and sort of ephemeral. And I think we need to develop as a community a better way of doing this. It's so kind of out there. But maybe that's beautiful, but it just seems like not right. Wes Bos I've redone my own custom, you know, spec flow like eight times. Yeah, so there's like some spec flow. Jared Palmer Anyway, figuring it out, but let's just put it in a markdown file. But then what I do is I ask it to create a rubric for grading the spec or this plan, then grade it, then revise it. Yeah. And that extra step of like it thinking through the rubric and then self-evaluating, I've found to be quite useful, but it can go overboard. And so it ends up like, I've probably seen this too. It's like week one to two, week three to four. I haven't figured out how to get it to stop planning like 12 weeks. Wes Bos What is that about? We did like a vibe code off. What is that? What is that? How do I get it out? Yeah. And then mine was like, We were seeing who could get it done the fastest. Like, oh, can we get this done in 10 minutes? Oh, my gosh. And then it was like, OK, week one. I'm like, no. Why does it do that? Scott Tolinski At the end where it tells you what it did, you got to tell it, stop yapping. Stop yapping. Yeah. Just give me the code. Jared Palmer So funny. But yeah, so if I could figure out how to not have it make a 12-week plan, that would be great. And I'm too lazy to correct it. Yeah, no, you're just like, yeah, yeah. Maybe it's good. I don't know. So that's some of my tips. The rubric thing is helpful. Ultra thing is helpful. ([Time 0:13:31](https://share.snipd.com/snip/c31a1fa5-744e-42ca-8356-c3dd9692c632)) - **Hedge With Parallel Prompt Variants** - Run multiple prompt variants or parallel attempts to hedge outcomes and improve chances of a useful result. - Jared suggests firing several model runs or using work trees to compare variants quickly. Transcript: Jared Palmer Something the cursor is going to have this very soon too which is the ability to have like n attempts or variants at the same prompt oh yeah just you know throw throw money yeah yeah yeah Just run them all away and like like branches or something like that i think they yeah yeah work trees um i think that's probably i mean yeah like let's just you know yeah ([Time 0:15:52](https://share.snipd.com/snip/940d168b-205c-47a1-b1b7-a6a742b735dd))