Transcription
Recently, people started talking about the AI bubble, saying that it was going to burst anytime soon with the disastrous launch of GPT5. Be that as it may, if the last model that we're getting is Claude Sonnet 4.5, then that's a win in my book.
Anthropic has outdone themselves and released the best coding model in the world. I mean, they do say it every time. But to be fair, they also deliver on that promise. Along with Sonnet 4.5, they've also switched the version number on Claude Code to two now. And this time they actually gave us features that someone asked for. Long-running tasks is one of the biggest improvements. Maintaining focus for more than 30 hours on complex multi-step tasks. They presented this case in which it coded a fully functioning Slack clone with 11K lines of code for 30 hours straight. Now whether this claim is true or not, the model does offer improvement in long-running tasks.
And from this, I thought of two really amazing use cases for this, in which Claude is going to extend from just being a coding agent in your terminal. But a $200 invoice for Claude Code always burns a hole through my wallet. So, here's a word from our sponsor, Ready, the AI website builder that makes your sites not just functional, but stand out. If you're building a single-page portfolio or landing page, Ready nails it with clean, responsive layouts you can instantly download and run locally. And with the selector tool, you can refine any section on the spot. No messy edits, just precise control. But Ready doesn't stop there. For multi-page apps, it generates reusable components, keeps navigation consistent, and even fixed a missing hamburger menu automatically, almost like it knew what was needed. That adaptive output shows real intelligence in design. Where truly shines is full-stack support. Unlike most AI tools stuck on the front end, Ready builds real back-end routes in Next.js. Need a working contact form API? Done. With stack selection, reusable components, back-end integration, and sharable previews. Ready is a true full-stack AI builder. Click the link in the description and start building today.
The first use case is centered around Claude Code becoming a general agent. I want you to take a look at this agent. It's an AI model release monitor. What it does is monitor the websites of all these big model providers and it automatically runs twice a day. It then sends me a Slack notification. Usually, to build or set up a production agent, you either have to go to another third-party service or build and deploy one yourself, which is way too technical. But Claude Code has become so good as a general agent that it's not only for coding anymore. It can virtually be used for anything.
Claude Code could have done this automatically in the terminal, but it wouldn't be automated. You wouldn't get notifications automatically. You'd have to open up your computer and type everything in the terminal. What I've done instead is use GitHub Actions, which you can think of as an automation hosted on GitHub. I've hosted Claude Code on GitHub using the Claude Code app, which is automatically provided by GitHub. To initialize the GitHub app inside your repository, all you have to do is run this /command. It's a guided workflow. It installs the app into your repo. Once that is done, you just have to prompt Claude Code with what you want in the GitHub workflow, and it'll set it up. But at the start, I faced true hallucination hell, as Claude made a lot of errors. From all of this, I made two rules for making automations on GitHub, and I'll explain them along the way.
Now, I told it exactly what I wanted to build. If I head over to Slack, you can see that we have this news update here from a bot we added. It automatically pulled important news from all these model providers, which was exactly what I wanted. Apart from the automated runs twice a day, you can also execute this manually. In the Actions tab, you'll see I have this AI model release monitor. The last two are the default ones added only for coding purposes. The second use case I'll be talking about is related to coding, but we'll get to that later. If I go into its runs, you can see nine total runs. It only worked correctly on the last one. To avoid this pain, you need to follow the rules.
The first one is that you need to tell Claude to auto-approve all the tools you plan to use in a GitHub workflow. I lost a lot of time because the culprit was simply that tools hadn't been auto-approved. Since this is an automated workflow, no one was there to approve them manually. Another thing that makes this really powerful is that you can add MCPs, and your automation can do stuff on other platforms as well. But whenever you want to add external MCPs, you need to test that the MCP servers are working correctly. Even if Claude fetches them online, you need to make sure they actually work. For example, we initially had a problem with Slack configuration. Claude thought Slack supported web hook URLs directly, but in reality, it required creating a whole app. The fix was just using curl commands to post directly to the web hook. That was one problem, and the other, as I mentioned earlier, was getting tool permissions right. The final fix was adding the allowed tools flag, which made sure Claude could run automatically.
Now, if I start the workflow, it sends the message in Slack again, and this time it just tells me there is no news. By the way, if you're curious about the genuinely useful new features in Claude Code, then first is the rewind feature, which restores the conversation to a previous point. In the earlier version, pressing escape twice would take you back to an earlier prompt. Now, when you rewind, it asks whether you want to restore the code as well, meaning it will reverse the code changes, just like we had in Cursor. The second standout feature is the improved IDE extension. It looks great now and is far more usable than the standard terminal version. There's still a minor transparency bug, but aside from that, it feels solid. You'll find plan mode and several other options, plus the ability to select slash commands from the menu, manage modes, and see which files are in context. Overall, it's much more usable with a cleaner, more intuitive interface.
Moving on to the next use case. This one is heavily inspired by OpenAI Codex, which provides you with cloud-based agents. You can give it a task, and it will automatically implement it in your GitHub repository. Since Claude's new model is specialized in long-running tasks, this was a really good use case to hand off background tasks. Even if you're using the BMAD method, you could tell it to implement a story, and it would just go ahead and do it. The best part is you can even run all of this from your mobile phone. That's the main goal of this use case.
To demonstrate, I've made a small demo app. It's a file system app that I've pushed to my GitHub repo. You can see the login page is already there, but no login functionality is actually implemented. So Claude's going to implement that using our cloud-based system. What makes this unique is the ability to add MCP servers. The authentication will be implemented on Supabase. The agent will handle it, but it would still need a lot of API keys from Supabase. But we're going to automate it all with the Supabase MCP. Let me show you the implementation. Then I'll explain more generally how you can connect any MCP server because I faced quite a few errors along the way.
The first thing you need to do, once again, is install the GitHub app. You don't need Claude to modify anything unless you want to add an MCP server. If you do, then there are a couple of additional steps I'll cover later. If you want to open a new task, think of it like opening a new issue. You open an issue and give it a title. I asked it to add authentication and wrote a short brief of what needed to be done. Then you tag Claude. It reacts with this emoji, which means it has seen your request and will start working. The error I ran into was that even though I first added the MCP, the tools weren't available. This goes back to the earlier problem. You always need to tell Claude to auto-approve all the tools you plan to use. And you also need to test the MCP servers. Since I didn't know that at first, I ran into both problems. Plus one more.
Claude did implement a base in the code, but it stopped because it needed the API keys for the complete setup. I told it to use only the Supabase MCP, but it got stuck since the tools weren't there. I tried prompting it repeatedly, but it still couldn't finish because Claude Code didn't have the context on how to configure the GitHub app. Now, the GitHub app is based on the Agent SDK from Anthropic. Claude could figure most things out, but it needed either the TypeScript or Python SDK documentation to complete the setup. You'll see a copy page button in the docs. Copy that and give it to Claude Code. That solved the errors for me because otherwise, it simply didn't know how to use the SDK. That was the first big fix. Then, coming back to the earlier rules, I didn't tell it to auto-approve the MCP tools. Another mistake I made was using an old Supabase MCP without testing it. That MCP only allowed read conditions instead of full read/write functionality. So I had to replace it with the correct MCP.
After resolving those issues, Claude was finally able to complete Supabase authentication using the MCP tools. It confirmed the implementation was successful right after I provided the SDK documentation. That step is really important whenever you're working with or building new custom agents. Always provide the documentation. Afterwards, it asked me to create a test user. I did, and then switched to the authentication branch. Now, if I open the app, you can see that the authentication is fully implemented. I logged in with the test user I created, and it worked. Authentication is now up and running, and the feature is implemented.
That brings us to the end of this video. If you'd like to support the channel and help us keep making videos like this, you can do so by using the Super Thanks button below. As always, thank you for watching, and I'll see you in the next one.