Woah, today’s reading list seems almost entirely focused on advice and good practices. I like when that happens. Invest in reading a few of these!
[blog] Building Etsy Buyer Profiles with LLMs. Etsy cares a lot about personalization, which makes sense given their business. Here’s how they built their new experience.
[article] Stop Force-Feeding AI to Your Developers. I actually think you SHOULD force-feed AI to developers, but I agree with this post that you shouldn’t use it to mask over other productivity investments that are needed first.
[blog] Serving Gemma 3 on GKE with TPUs and vLLM. Use GPUs or TPUs, either are fine with me. It’s great that you can use either in a Google Kubernetes Engine cluster.
[blog] Claude Code: Now in Beta in Zed. This fast code editor got their Agent Client Protocol (ACP) rolling last week with us and the Gemini CLI. Now they continue bringing agent options in with Anthropic’s offering.
Heading up to Mountain View to participate in some customer meetings at the mothership. It gave me an excuse to update some demos, and even prep some new things.
[blog] Running Secure Kubernetes (GKE) Workloads in GCP. This is the type of content you want loaded into your LLM context so that you get reasonable advice when you ask to “build a checklist to secure my Kubernetes cluster.”
[article] Is Your Talent the Bottleneck to GenAI Success? I’d suggest that in almost every case, the answer is no. You actually have the talent in place already. What you may not have done is properly train them and invest in their skills.
Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:
Happy Fake Monday to my American readers who stumbled through a Tuesday after a long holiday weekend. The reading list today has some industry news, and also a handful of pieces that offer great insight into AI topics.
[blog] 📖 An Open Book: Evaluating AI Agents with ADK. Very good post that clearly explains how to create eval sets and test criteria before executing agent evaluations using the Agent Development Kit.
[blog] Go 1.25: The Container-Native Release. Go is already great for container workloads, but it looks like this latest release added some major improvements.
[article] Anthropic raises $13B Series F at $183B valuation. The spigot for AI funding is still flowing, and it’s going to be tough for new entrants who need to commit to a LOT of compute to keep up.
[blog] Mass Intelligence. When you think about it, the fact that nearly everyone on the planet has access to unprecedented “intelligence” is bonkers. Ethan shares how we got here.
[article] From Black Box to Blueprint. Can you reverse engineer a system with the help of AI and use that insight to build a modernization blueprint? Thoughtworks shares their experience doing just that.
[blog] Do the simplest thing that could possibly work. I think we all know this, but it’s hard to do. We want to predict the future and build solutions, platforms, and structures that suit our expectations of what’s coming next.
[article] AI Contrarians on the Problems With Vibe Coding. It’s ok to be contrarian. I’ve been using these tools a fair bit, and have learned that it’s not all or nothing. With the right expectations, and using the right tool for the job, it’s objectively better.
[article] Legacy tech is hard to kill. Yes it is. Much easier to just add new fancy stop on top and leave the lower layers for the next sucker.
[blog] No More Coders? You Still Need DevOps. Andrew doesn’t believe that coders are going away, but execs who are hedging their bets might be de-emphasizing platform investments. Don’t do that.
I’m often grateful for all the individual contributors I work with, but today I was extra thankful to observe so many good managers in action. Management is often a thankless task. But the good ones expertly advocate for their staff, set smart priorities, and lead by example. We’ve got that in spades.
[blog] Some thoughts on LLMs and Software Development. Scattering of thoughts, but those who pay attention to many dimensions of our industry often have genuinely interesting perspectives about AI.
[article] LLM System Design and Model Selection. Nice writeup with tons of questions and dimensions to consider when choosing an LLM and LLM-based system architecture.
Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:
My favorite movie is Tommy Boy. I apologize for nothing. You won’t find this flick listed on the top 100 movies of all time, and some people think it’s a dumb movie. That’s ok. Look, I don’t really get the Dave Matthews Band. They seem like nice chaps, but it’s not my thing. But they have a massive following of superfans. Everything isn’t for everyone. Not every product is built for you.
I recently looked at which AI coding tool was the right fit for a given situation. But what about Google’s hefty portfolio of products for those who want to vibe code and let AI take the wheel? In that case, it’s not just about the given situation, but also the type of person. Not every product is for each type of user.
Vibe coding is expanding the pool of people who build apps. It seems to consist of consumers who are non-tech folks who want to bring ideas to life. There are tech-adjacent professionals who do “knowledge work” and might be business analysts, product managers, program leads, and executives. And then you have software developers who have deep understanding of tech, and want to quickly produce new software.
What are we vibe-coding? To me, it seems like we’re building throwaway prototypes to just try something out quickly. We’re creating personal software that’s meant to improve our own productivity. And we’re delivering multi-user apps that are intended for legit use by others. So in my mind, it’s a grid like this, with my take on which Google tech fits where:
Throw-away prototypes
Personal software
Multi-user apps
Consumers
n/a
Gemini Canvas Gemini Gems
Gemini Canvas
Tech-adjacent professionals
Gemini Canvas Google AI Studio
Opal Google AI Studio Agentspace
Gemini Canvas Google AI Studio Opal Firebase Studio
Software developers
Google AI Studio Gemini CLI
Gemini CLI Gemini Code Assist
Gemini CLI Gemini Code Assist Jules
Vibing as consumers
I’m not sure consumers are interested in throwaway prototypes. My non-tech friends wouldn’t want to geek out on tech. They have some sort of goal to solve a problem.
What about consumers building “personal software” that acts as a web app, agent, or tool? Sure. Gemini Canvas seems like a good choice for this. And, for building simple apps to share with others. Gemini Gems are a tool for building personal AI assistants without needing to be a tech expert. Some are creating fun consumer-grade demos with Google AI Studio, so I wouldn’t complain if you added that product into this row as well.
Let’s look at Gemini Canvas. You activate this in Gemini when you choose to “build.”
Let me provide it a simple prompt:
Beautiful exercise tracking app that lets me record the exercises I did in a given day, and for how long. I can also view past days and see trends over time.
When I enter that prompt, Gemini gets to work. It creates a single-file app where the code is not the focus. You can see the code, but it quickly switches the UI to a preview of the app. I can make changes via the chat and get a live look at the changes.
This is a fun to use, simple interface that’s consumer friendly. It’s easy to trigger, very fast at generating apps, has basic abilities to rollback changes, and offers sharing via a public link. It’s very opinionated on the tech stack, the code is all stuffed into a single artifact, and you don’t get many legit deployment options. Great for consumers who are building personal software or simple apps for a small group to use.
Vibing as tech-adjacent professionals
I think it’s awesome that anyone within a company can be a builder. It doesn’t matter if the HR person, executive assistant, program manager, or VP doesn’t know how to code. They can use Gemini Canvas as I showed above, along with other tools.
Some who have some tech familiarity might jump to Google AI Studio. It’s free to use and fantastic for builders. From the “build” menu, you can trigger a vibe coding experience that keeps the focus on the outcome, not the code.
I’ll use the same prompt as above, but you also get a few other configuration options, including the ability to choose between React or Angular.
Once I submit the prompt, Google AI Studio gets to work thinking through a plan and building out the components. The interface is terrific here. What’s different from Gemini Canvas is that you get a professionally structured project with code arranged across files. It takes longer to get to a Preview because it’s doing more (responsible) work, but it’s still very fast.
I like the suggestions offered for the app (above chat box), easy ability to download the app, GitHub integrations, and one-click deploys to Google Cloud Run. The code editor is basic, so I wouldn’t use this for sophisticated builds, but that’s not what it’s for.
Opal is a new Google experiment for building “mini-AI apps” and isn’t a standard vibe coding tool. Think of it as a way to build apps that are focused on generating content with AI.
My exercise tracking tool doesn’t make a ton of sense here. One use case for Opal could be to generate text for sharing content on each social media site.
Opal lets you define what you need to collect from the user, assets you have available (YouTube videos, documents, and more), can perform web searches, generate all sorts of media, and aggregate results.
It’s simple to preview and share these apps, and I like the concept. It’s not a traditional “vibe coding” tool, but I can see where non-developers would like using it to bring ideas to life.
You could also consider Agentspace a vibing tool for office workers. Agentspace is a unique platform for those who want a more useful and effective internal experience for getting work done. Besides offering a AI-enabled search and research, it also has an agent-building experience for those who want personal agents.
The agent builder interface is entirely no-code, and lets you tap into public internet searches, along with private enterprise data sources and tools. This is super helpful for those who want to automate repeatable tasks or build personal productivity solutions.
The final option I put into this row is Firebase Studio. This is a service that’s completely applicable to software developers, but also friendly to those who aren’t professionals in this space. Unlike most of the options I’ve listed so far, this isn’t only for front-end solutions. I can build backend Go or Java apps too. It also offers a vibing UI where you can start with a prompt and build the app. I’ll use the same prompt I did earlier.
When you start vibe coding here, Firebase Studio shares an app blueprint and then gets to work. I wouldn’t give this to a consumer persona—there’s still technical know-how you’d need to have—built it’s approachable to those who aren’t full-on software developers.
Vibing as software developers
If you’re a developer, you can use any of the options above. You might love the simplicity of Gemini Canvas, or prefer the opinionated Google AI Studio environment. Sounds good, live your life.
Many software people want to vibe code with tools already in their toolchain. Your best bets with Google are the Gemini CLI and Gemini Code Assist.
The Gemini CLI is a command line interface that’s free to use. You can authenticate with your Google ID (as an individual or corporate user), bring a Google AI Studio API key, or use Google Cloud Vertex AI. It has built-in tools (Google Search, shell, reading files, etc), supports MCP, has configurable memory, and can run anywhere. We recently added a supporting GitHub Action so that you can use it in your code repo. And Zed just integrated it into their next-gen code editor.
The Gemini CLI is a great vibe coding tool. I could use the same prompt above, but also ask for a technical spec first, define my language/framework preferences, and generally steer the build the way I want.
When I want maximum power during my vibe coding sessions (like I had yesterday), i use a combination of Gemini Code Assist with the Gemini CLI rolled in. Killer combo, as I get the conversational AI mode of Gemini Code Assist in my IDE/editor, but also the rich agentic power of the Gemini CLI with shared context. Yesterday I vibe coded a complete “travel app” after using the CLI to generate a spec and then incrementally implementing it with the CLI, and taking over control in the editor when I needed to.
Wrap up
There’s no wrong answer here. Use what fits your situation, and the role you’re playing. If you’re goofing around and just want a frontend app, anyone should use things like Google AI Studio. If you only want a code-centric experience for every situation, stay with IDE-style tools and CLIs. But I love that it’s so much simpler for people of any skill level to realize their ideas through software thanks to a range of vibe coding tools that suit each person’s taste.
I vibe coded an app today that uses our fancy new image generation model. My customer presentations are more interesting now (I think!) when I can create personalized apps and content for a customer.
[article] Google’s Gemini CLI Agent Comes to Zed. I’m running three different code editors now, which would have been absurd just a couple of years ago. Zed is pretty great, and now can incorporate agents like the Gemini CLI. Our post and Zed’s post cover it too.
[article] IT, business leaders clash over cloud, data security. Maybe that dissatisfaction with cloud/AI/automation can be found later in the article. Supporting technology and security approaches don’t seem to be ready at these companies.
[blog] APIs don’t make good MCP tools. I’m persuaded by these points. Just handing a pile of APIs to an agent doesn’t seem like an efficient thing to do.
[blog] How to Fix Your Context. This is from a couple months back, but I don’t think I saw it then. Or if I did, and had it in a reading list, I forgot about it. Either way, a helpful read.
One reason that I could never be a solo worker is that I like the epiphanies that come from conversations with smart colleagues. Today there were a couple, and it makes me better at what I do.
[blog] Who are we designing for now? You’ve got to ask and answer this question. Your UI and APIs now have agentic consumers which require new patterns.
[article] Developers lose focus 1,200 times a day — how MCP could change that. I’m down with this. We’re all using too many apps and context switching more than needed. If MCP (or whatever) makes it easier to pull the data and actions we need into the place we’re at, it’s a win.
Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:
Tonight I’m attending an in-person draft for the Fantasy Football league I’m in. I asked Gemini Deep Research for a draft strategy given the parameters of the league, so this may be how I’ll know if we’ve achieved AGI or not.
[article] The Number One Productivity Killer for Devs? Finding Information! Some dire findings here. Developers can’t seem to find what they’re looking for, even though AI tools are saving them 10 hours a week. And, the mismatch with executive expectations continues.
[article] Does AI spell the death of front-end engineering? Death? No. But I do wonder about the impact now that it’s much simpler for a backend/fullstack engineer to craft functional frontends with AI.
[article] Outdated Python Versions Cost Companies Millions. Even if you don’t care about the features in a new version of your programming language, you’re likely missing out on key performance updates that can save you real money.
[blog] A Developer’s Guide to Model Routing. Learn why model routing is important, and then see a good example of building your own lightweight router.
[article] Why getting to bug zero is so hard. It’s not so simple to get rid of all the bugs in your issue list. Tom looks at a spectrum of bugs and how to tackle them.
[blog] The Story of Valkey. It’s an interesting story, and one that’s showing surprising staying power.
[blog] Gemini CLI Observability with OpenTelemetry. Cool post. I didn’t know about (or, forgot about) the OpenTelemetry support for our agentic CLI. I’d imagine that getting metrics about token and tool use could be very important.
[blog] Build vs Buy in the Age of AI. I liked the perspectives here on “user programming tools” and why business software won’t get replaced by vibe-coded alternatives, but big changes are still coming.
Want to get this update sent to you every day? Subscribe to my RSS feed or subscribe via email below:
[blog] Caching Strategies for AI Agent Traffic. Caching is caching, but the patterns you apply can evolve. And it looks like AI agents will force us to shift our thinking.
Richard Seroter's Architecture Musings 