Stable Discussion Podcast

About Stable Discussion Podcast

Artificial Intelligence is changing our world and we help better understand what this means to all of us. We'll look at what's possible and where is the technology still not there yet. blog.stablediscussion.com (https://blog.stablediscussion.com?utm_medium=podcast)

Publisher

What's possible with AI today and what to expect tomorrow

Category

news · technology

Language

en

Explicit

No

First episode

30 Jan 2023

Latest episode

19 Nov 2025

Latest episodes

3 of 35 episodes have a transcript.

1. 

   19 Nov 2025

   Google's Antigravity IDE - Still too Early

   I played around with the new Google Agentic IDE, Antigravity (https://antigravity.google), on launch day and created a few features for an app I’ve been playing with. If you’re unfamiliar this video is a helpful overview of the features: My initial impressions offer a more nuanced experience than the chipper attitude of this presentation which should help you get a balanced perspective. This Substack is reader-supported. To receive new posts and support our work, consider becoming a free or paid subscriber. Using Antigravity Agent Manager (https://antigravity.google/docs/agent-manager) This interface feels like a move in the right direction. It offers a means of managing the work done by an agent, ability to see and respond to plans easily, and clear indication of changes made. I like the Agent Manager’s UI but it’s been a little buggy so far. I made some good changes but it is missing some of the context I have in my CLAUDE.md files on how I wanted it to build the app. I’m not sure if it’s reading some of my core information files or docs. Knowledge (https://antigravity.google/docs/knowledge) The knowledge base feature looks interesting but in the two medium-sized features I created it didn’t seem to think this needed to be updated. Unsure when it will feel there’s something worth of it. As with all AI memory systems I do worry about it getting the wrong idea and storing that idea for later use. Intelligent Tool Approval (https://antigravity.google/docs/agent-modes-settings) Antigravity lets the model choose when to run a tool in some cases rather than stopping to ask for permission. This is a cool concept if it works well and I’m curious to explore with this further. I worry that seemingly innocent commands may look non-destructive and get called anyway by the model despite causing some destructive change. This may require a metadata layer similar to how MCP servers have expanded their interfaces to include if a tool mutates data. Browser Tool (https://antigravity.google/docs/browser) It’s great to see them add a tool for interacting with the code as it’s actually run. However, my initial setup of their browser didn’t seem to workout very well. I installed the extension but the agent had difficulty finding the extension. Eventually it seemed to work when I opened up a new tab after closing Chrome. Since Chrome isn’t my main browser, it’s not setup quite right for the application I was testing but it worked well on a later project. It seems to be able to record, capture screenshots, and read the console. Commenting (https://antigravity.google/docs/implementation-plan) Having a highlight and comment system on AI plans is plain great UX. In existing solutions I find that there are often times where I’m opening a note or something to put my feedback in and then scrolling down and pasting that feedback into the chat input at the bottom of a agent chat window. When I made comments, it seemed like they were factored in appropriately when I asked the model to apply them. The pitch of this feature sounded good but I’m curious how the model thinks about incorporating feedback like this in practice. Gemini 3 Pro (High) I’ll pay it a good compliment in saying that this model felt a lot like Claude Sonnet 4.5 to the point that I felt like it was working with me in the same way I’m used to. It’s not often that happens when switching models. That said, I still didn’t get my context appropriately in the conversation and worry a bit if I have to start all of my conversations with it saying Read @CLAUDE.md before we can start working. On release day there’s always a lot of strain on these models and this was no different. There were a couple times I needed to leave and come back when working through the features to let the global limit cool off a bit. Hopefully the usage gets a bit more predictable and lets more people use this regularly. Onboarding When I launched Antigravity and onboarded, the onboarding crashed on the final step. Then I had to go through it again and even though I had indicated I wanted it to pull in settings from my Cursor, it ignored that and didn’t pull in any of my extensions or config. This is definitely a headwind in my adoption journey but something that is likely to get fixed with time. Takeaways I built out two features at the same time to explore the agentic capabilities of the Antigravity IDE. One feature was to add notes for users in a feed and the other is to allow users to upload images to Google Cloud Storage using a rich text field. However, I think having the changes happening in parallel was a bad idea. Part of me assumed these changes would exist in separate worktrees or branches so that they wouldn’t conflict. Some of the demo videos made that seem like it might be the case but no, it’s the same as running two separate Claude Code instances in the same repo. Just a new UI. Ultimately, I wanted to put Antigravity through its paces, but running two things at once confused me a bit while learning a new tool. It also seems to have confused the interface too because one of the agents just stopped responding to my prompts after it tried unsuccessfully to test the feature in the browser. The other agent completed it’s work fine but in the Review tab was still showing both sets of changes which was, again, confusing. At the end of my coding session I had problems determining how to progress and what to do with the resulting conversations. There’s an ability to review the changes and provide feedback but it’s still a bit confusing how to get the IDE to commit the change from the Agent Manager view. When I did merge I also didn’t know what to do with the conversations. I wish there was an archive feature or something as deleting these conversations doesn’t feel great especially when the Knowledge doesn’t seem to update. To complete my changes, I ended up just switching back over to Claude Code as that seemed to have overall better context on what I was building and I had better muscle-memory as to how to progress a late stage change. In Antigravity, there’s a lot of really great intentionality around context but some of the control around context does feel limited. Because it’s a “smart system” there’s a lot less control. That’s helpful in someways but also makes it more difficult to understand exactly what’s going on at any time. I keep noticing some things I’ve come to appreciate about other interfaces that are missing here. One example is queued changes in Claude Code. If there’s a string of commands that make sense to just run one after another, it’ll queue them up. I find that while the auto approval works well in Antigravity, but I find there are times where I need to wait to approve several changes that were clearly known in advance that could have been approved without delays between each approval. Release Article & Videos The release blog mentions “Gemini 3 is also much better at figuring out the context and intent behind your request (https://blog.google/products/gemini/gemini-3/#note-from-ceo)” but I haven’t found this to be the case. Jumping into an existing codebase some of the core NextJS architecture I had in place for version 16 was ignored despite clear indications. That said, many of the solutions create were well done. They just didn’t retain the high context as this blog post might indicate. In the getting started video, it was refreshing to see the a Google engineer directly trust the AI with his API key. I think that’s honestly the norm in a lot of cases depending on how permissive the keys are. It enabled the AI to explore and investigate the API with context of the interface due to some Googling of the interface from the web. In all honesty, this is a pretty likely use case for most devs. Agents Testing During Research It’s amazing to see the impact of the Intelligent Tool Approval when it comes to agents doing research. That’s where there’s a bit of magic in this release. This makes me think that this might be one of the better agentic interfaces to do coding research within. Nano Banana Image Gen It’s awesome to have an image generation module as powerful as Nano Banana (https://aistudio.google.com/models/gemini-2-5-flash-image) running directly in the IDE. It can generate assets and directly add them to the application. That’s pretty incredible. (no transparent backgrounds though unfortunately) Next I’m intrigued by Antigravity but it’s still feeling a lot like a Beta of something that could be cool. It’s going to be interesting to watch other competitors in the space learn from these solutions and find ways to improve their services based on these changes. I wouldn’t recommend rolling Antigravity as your main editor for a few weeks while bugs get ironed out but I think it’s great to experiment with and potentially run research tasks on existing projects. This Substack is reader-supported. To receive new posts and support our work, consider becoming a free or paid subscriber. Get full access to Stable Discussion at blog.stablediscussion.com/subscribe (https://blog.stablediscussion.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)

2. 

   20 Oct 2025TRANSCRIPT

   AI-Assisted Coding Features with Claude Code

   In my latest video, I share a high-level summary of building a full feature with Claude Code and AI-assisted coding — from rapid prototyping to a frustrating production bug that ate up 20% of my time. If you haven’t seen it yet, you can catch the overview here: Videos are great for telling the story at a glance focusing on the high-level summary — but they don’t always have the space to unpack the details that really outline the feel of building something with an AI. That’s what this publication is for. Unpacking the hype and diving deep. If you’re building with AI tools or just curious about the real-world bumps on the road, this deeper dive will give you practical insights and a more nuanced perspective. Lessons Learned in the Coding Session I’ve captured all of the 88 messages I sent in this coding session in one consistent webpage (https://claude-code-session-messages-displa.vercel.app/) where they can all be browsed at your convenience. There’s also this companion site (https://project-coding-session-dashboard-874.magicpatterns.app/)where I outline some of the high level takeaways. In this post I’ve pulled out a few gems that highlight some tips and tricks that I think are pretty critical to having a good experience doing AI Assisted Coding. For each note, you’ll notice the note number (message number), a quoted text that indicates the message I sent the AI, and a description that details what I was doing at each step. For example, the first major message I sent was this one: Here we’re catching Claude Code up to some of the research I did to begin the feature. You’ll see the comment there and a quick note with the details. This is everything I needed to give the AI to get the feature started. Let’s look at a few other notes that are more telling Providing Context7 to Documentation Context is super important to get right when working with AI, so much so that there’s an entire Context Engineering movement. As such, I often leverage Context7 (https://context7.com/) to pull docs from technologies I’m coding with to leverage in my code. Today’s AI models are trained at some point in the past and all contain a timestamped view of the world and pulling this context ensures we’re aligned with the latest version of any technology we’re using. This can have a huge impact on how well a feature is integrated in our code. I’m using this to ensure we’re correctly calling the API and ensuring that we’re downloading the image from that API in the right format. Note: I’m also not generating code right away. I’m using this research to put together a document that we can reference later when we build out the feature. I find this helpful to ensure we keep available key technical considerations as Compaction in our conversation (generated summaries of conversations) may cause some details to get lost. Prompt to Prompt As I’m building features that, themselves, leverage Large Language Models, I’m needing to create some prompts and tweak them as I go. I generally will tweak a bit by hand but I’ve found as I continue working with Claude Code that it’s often more explicit and creative at coming up with good prompts. Anthropic themselves also have a really great prompt creation tool on their Claude Console page (https://platform.claude.com/login?returnTo=%2F%3F) for developers to take their prompts and refine them into more optimized prompts for working with their models. Generate Test Data I’ve found AI uniquely helpful in getting some test data spun up quickly for testing out features. If you let it go a bit wild you’ll sometimes be rewarded with something cool or unique! Try it out sometime! Regular Planning is Key Using Claude Code, planning is a pretty critical step. I will generally add planning in as a step whenever my direction shifts fundamentally. Here I’ve realized I have a new page that I need to develop that leverages the feature we’re working on but will require a much wider sense of the application to be able to pull off well. At these times I want Claude to be investigative and curious rather than ambitious and assuming. If I just let it go on a feature like this it could easily create a page I don’t want or add unrelated changes that don’t line up with the direction I’m moving towards. Give Claude Code Your Tedious and Hungry I love tasking Claude Code with fixing the code around the changes that have occurred. When I use AI Assisted Coding tools I also couple those tools with a suite of analysis tools to test the code that is generated for common errors and mistakes. This is key to going fast and projects that I don’t have setup well in this space move slowly even when I use AI Assisted coding practices. It’s plain painful to get them moving as I need to be much more diligent as to what changed and why. That said, you may need to poke it a few times to do what you want. AI will often take shortcuts in order to try to focus on what it sees as it’s main goals and priorities. This means occasionally you’ll need to redirect those goals a bit to align with your own. Regularly Document Wins When I finally pushed through the major defect that had come up, outlined in the video, I realized we needed to update our documents to capture this. It wasn’t captured anywhere else yet and it is critical to get this captured so we can refer to the fix in the future. An Image is worth a lot of Words Passing images to Claude Code is an amazing way to tell it exactly what you have in mind. I regularly pass images to it when I’m working to have it understand something a bit more complex or need to simply point at something for it to understand. Reference images can help a lot when coming up with a design or theming around a color. I will regularly leverage a visual vibe coding tool like Magic Patterns (https://www.magicpatterns.com?via=ben-hofferber) to build out ideas for me to quickly reference via screenshots or code snippets. I’ve got a bit more about how I use that tool here: The Overall Journey Building with Claude Code is exhilarating because you can go from idea to implementation so fast. But as I’ve learned, it’s just as important to slow down sometimes — to observe, orient, and decide — before you act (https://rangle.io/blog/fix-vibe-coding-with-ooda-loop). AI can supercharge your coding, but it can’t replace the human insight that keeps your code cohesive and aligned with your goals. Going back through these messages has been a great way to surface some of the ways that I work with Claude Code to prototype. As you go through them, I’d love to know if you found anything interesting about the way that I work with Claude Code that I didn’t mention. Additionally, I’d love to hear how you’re working with AI-assisted coding tools. What bumps have you hit? What tricks have you found? Drop a comment or reach out — let’s keep unpacking the hype and learning together. This Substack is reader-supported. To receive new posts and support our work, consider becoming a free or paid subscriber. Get full access to Stable Discussion at blog.stablediscussion.com/subscribe (https://blog.stablediscussion.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)

3. 

   8 Oct 2025

   Thoughts on OpenAI's Dev Day 2025

   I wanted to collect a few thoughts on the recent OpenAI Dev Day 2025 announcements from my initial investigation into the tech behind the announcement. Based on a couple years of building AI integrations into applications these are my gut reactions to the presentation.If you’re unfamiliar with their presentation, it might be valuable to skim through it first. Now let’s get into it. Our Take AgentKit (https://openai.com/index/introducing-agentkit/) and Agent Builder feels great and looks like what the future of tools for building Agents around real products and services will look like. It’s like an extension on n8n’s existing capabilities (another agent builder tool) but AgentKit is streamlined around OpenAI’s offering. This could be serendipitous if you’re already leveraging their existing file storage solution or other core features. The kit is streamlined for building quickly but I don’t really think it’s quite the powerhouse people think it is. We can see why by reviewing some of the other features in the announcement. ChatKit (https://platform.openai.com/docs/guides/chatkit#add-chatkit-to-your-frontend) is an application layer toolkit for delivering a chat interface to end users by directly leveraging ChatGPT’s methodology. It has a nice set of features that manages to do many of the things that the Vercel AI SDK (https://ai-sdk.dev/docs/introduction) was already doing well. I’m a fan of some of their direction which is clearly inspired by that library. Similar to building with Swift for the Apple iOS, to leverage this kit a team will be aligning with a design system of visual components created by OpenAI. Until we get a chance to play with these components (https://widgets.chatkit.studio/components/button), it’s unclear how far these components can be pushed and where the limitations are. Additionally, we’ll need to wait and see how this library will evolve over time since this product space is certainly very new. OpenAI is looking to own more of the experience layer by providing an ecosystem of UX and UI tooling. Applications leveraging their agent platform will need to keep pace with changes and adjust accordingly if they adopt this approach to building. That can be an uncomfortable place to be longterm and might be a worry for early adopters. OpenAI mentioned that there would be a capability to publish Apps in ChatGPT (https://openai.com/index/introducing-apps-in-chatgpt/) in the future but no word yet on exactly when. The actual guidelines (https://developers.openai.com/apps-sdk/concepts/design-guidelines#best-practices) to publish an app are quite extensive however. They remind me of an Apple App Store-like approval process blocking publishing and becoming featured. Adherence to the style and intent of ChatGPT will be directly rewarded here. AgentKit doesn’t directly land users into ChatGPT, which can be a bit misleading if you watched the presentation. It seems like AgentKit has everything setup to build something into ChatGPT itself but in actuality AgentKit is for creating agentic experiences on separate company-specific websites. As OpenAI leans into adopting MCP, there seems to be some underlying messaging that companies don’t have a vendor lock-in around OpenAI. However, the MCP ecosystem is still missing many core services and is still maturing. I’d argue there is a lot of inherent vendor locking with AgentKit. That much is clear. Evals is one of the most compelling reasons to be excited about AgentKit. But to leverage it well teams will need a very clear vision of what an agent does and what it looks like when it does something well. That continues to be a difficult spot for product builders to define. Overall, I think AgentKit shows an interesting perspective on what agentic platforms should look like. Unless there’s a clear path towards Apps in ChatGPT I think the main adopters of these releases are going to be B2B application builders. While there’s room for a B2C path, losing brand seems like it lacks competitiveness and limits the upside potential. Existing options for building agents continue to be available and, given a team with some frontend engineering capability, those solutions aren’t as complex as OpenAI makes it sound. Get full access to Stable Discussion at blog.stablediscussion.com/subscribe (https://blog.stablediscussion.com/subscribe?utm_medium=podcast&utm_campaign=CTA_4)

Inside a recent episode

AI-Assisted Coding Features with Claude Code

Published 20 Oct 2025 · Transcript excerpt

[…] For example, the first major message I sent was this one. See note two. Here we're catching clawed code up to some of the research I did to begin the feature. You'll see the comment there and a quick note with the details. This is everything I needed to give the AI to get the feature started. […]

Reach and audience

Public platform figures. Ratings count people who left a rating, not total listeners.

YouTube views

221,063

Score snapshot 15 Sept 2025

Spotify

5.0 / 5

3 ratings

Podcast Authority Score: 50 / 100

A composite of feed quality, social presence, YouTube performance and engagement. Read the methodology.

Quality

42

Social presence

0

YouTube

86

Engagement

60

Contact Stable Discussion Podcast

Guest appearances

Does not typically book guests

Based on episode analysis; this does not confirm that the show is currently accepting guests.