§ 01 · The Hook
The bait, then the rug-pull.
Brian Casel built Agent OS, one of the most-used Claude Code frameworks of 2025. Here, he opens by telling you it is mostly overkill. That self-subversion is the whole hook: the credibility of the framework author retiring his own work.
What the video promised.
stated at 00:55 "I want to show you what building with Claude Code and nothing else actually looks like here in 2026 and why the path forward is simpler than it looks." delivered at 09:28
Where the time goes.
00:00 – 05:09
01 · The 2025 noise problem
Frames the landscape: frameworks proliferated because models and tooling were immature. Establishes that 2026 is different. Introduces the live build premise.
05:09 – 08:23
02 · Plan mode + clarifying questions
Launches Claude Code in plan mode, voice dictation prompt, watches sub-agents explore codebase in parallel, reviews clarifying questions and plan.
08:23 – 09:28
03 · What changed in 2026
Two converging forces: models dramatically more capable (Opus 4.5), and Claude Code itself evolved with plan mode, sub-agents, skills, spec-driven workflow.
09:28 – 11:40
04 · Reviewing the build
Claude finishes Trends View. Brian reviews output, points out UI issues from not invoking design skill early enough, Claude fixes in one pass with screenshots.
11:40 – 12:53
05 · Core features that matter
Skills and context management highlighted. Craft argument: models implement patterns but cannot choose which pattern is right.
12:53 – 14:53
06 · When frameworks still help
Two remaining use cases: greenfield design (Design OS) and legacy codebases (slimmed Agent OS). Closes with subtraction-not-addition thesis.
Visual structure at a glance.
Named ideas worth stealing.
05:09
concept
Spec-Driven Development
Plan mode then clarifying questions then plan review then build. Now native to Claude Code, no framework needed.
Steal for Any agent-heavy feature build in JoeFlow or MCN+
00:45
concept
90 / 10 Rule
Vanilla Claude Code handles 90% of work. The 10%: greenfield design systems and legacy codebase conventions.
Steal for Positioning frame for MCN+ builder tools
13:51
concept
Subtraction Not Addition
Strip away unnecessary tooling. Only add complexity when you feel the friction that demands it.
Steal for JoeFlow positioning
Lines you could clip.
00:45
"Pure vanilla Claude Code with Opus 4.5 is all I actually need for 90% of my daily work."
11:40
"The bottleneck is not writing code anymore. It is knowing what to build and how to structure it."
12:20
"Your judgment, your taste, your product instincts. That is what makes it work."
13:51
"The mistake is assuming that you need extra tooling before you allow yourself the opportunity to be productive."
How they spent the runtime.
Hook length55s
Info densityhigh
Filler5%
Things they pointed at.
08:23productOpus 4.5
11:40toolClaude Skills (front-end)
12:53toolDesign OS
How they asked for the click.
14:13
subscribe
"Hit subscribe on the channel so you do not miss my next video when it comes out."
Teases follow-up on slimmed Agent OS — strong continuity hook before the CTA, no hard sell
Word for word.
HOOK opening / re-engagementCTA the pitch metaphor analogy
00:00HOOKClog code in 2026 is not what it was when it launched almost a year ago. And if you're just coming into AI assisted development right now, I get why it might feel overwhelming with all the noise of the past year. Frameworks and elaborate systems for managing your agents and custom cursor rules and MCP servers and swipe files for prompts. I mean, there's a out there telling you that you're using the tools all wrong or that you need to master this entire ecosystem before you can even be productive with Claude code. And look, I'm someone who created one of those frameworks myself, a popular one called agent OS. And I'll be the first to tell you that here in 2026,
00:36HOOKmost of it is overkill. These days, pure vanilla Claude code with Opus 4.5 is all I actually need for 90% of my daily work. Yeah. I said it. Now, before you think I'm about to tell you that my Agent OS project is obsolete now, hold that thought because in my next video, I'll show you the updates that are making Agent OS lighter and leaner for how we build in 2026.
01:00But here in this video, I wanna show you what building with Cloud Code and nothing else actually looks like here in 2026 and why the path forward is simpler than it looks. Now, if you're just finding my channel, I'm Brian Castle. I help professional builders stay ahead of the curve with AI. And every Friday, I send my builder briefing. That's a free five minute read where I give you my real take on what I'm seeing in our craft and our industry week to week. No hype, just what's actually working. You can get yours by going to buildermethods.com.
01:29And if you're serious about adopting Claude code this year, I just launched a new course for 2026 called build with Claude code and that's available to builder methods pro members along with our community of builders. And so to show you what I mean, I'm gonna build a real feature in one of my real apps right now. And so this is inbox summaries. It's a tool that summarizes feedback and messages from my audience and sends me regular reports and trends on and insights. And so this is the app that I have, uh, running locally here and, uh, and so here is like one of the reports that it sends me every week, um, you know, showing me some like audience research survey insight data. So I get these every week with some trends and some data points, But I wanna kind of pull together all of the overall insights into a new trends view over here. You know, something visual with like graphs and metrics.
02:17So I'm gonna show you how I would go about just beginning to plan this feature and then building out a spec and then we'll actually build it out in this video. So I'm gonna start by launching Claude and I have a shortcut to launch it in YOLO mode. I like how they still have the the the winter theme here. Now, first thing I'll do is tab into plan mode.
02:36Most big features, I always start off in plan mode. But I'm actually not even ready to have it create the actual plan or the spec just yet. First, I still need to plan out like what what's gonna go into this feature, what's in scope, what's out of scope. Right? So I'm gonna start off with a simple prompt and I'll just use some voice dictation for this. I want to add a trends view,
02:58analyze all of the functionality around those areas and suggest some ideas for what we might put into this trends view and make some suggestions as to which items
03:10should be included in a v one of this feature and which items might add complexity that we should probably push off for a future phase. So you know, I'm just starting off by voicing some of the actual strategic product planning questions that I would have in my mind and I would want to think through. I'm just going to pose those to Claude which has full access to my code base and and it's going to come back with some ideas. So let's let it cook on that. Now as you can see it's analyzing my existing code base and it launched three agents
03:41under the hood. You know Claude code just decided to do that. It's essentially using sub agents and they're running in parallel. It it sent off three different agents to explore, you know, different parts of my code base to help inform its answer to my strategic planning questions. You know, the amount of change that's happened just in the past year has been unbelievable.
04:01Less than twelve months ago, Claude Code was brand new. Yeah. The models back then were capable but inconsistent. We'd prompt and sometimes get great results and other times it'd be completely off. And there were real gaps in our workflow. And so those of us who've been excited about where AI in software development could go, who could see the potential even through the rough edges,
04:23we felt those gaps acutely. We wanted this to work. We could see it was close but the tooling wasn't quite there yet. So we built things to fill those gaps and that's why I created Agent OS. I wanted a way to inject my standards into Clog code. I wanted a planning phase that would ask me clarifying questions before building anything. I wanted task tracking and structured specs and I wasn't alone. Frameworks and custom setups proliferated last year because experienced builders were trying to bridge the gap between what Claude code could do at the time and what we needed it to do And that was the reality of 2025.
04:59All of that tooling created a kind of noise. If you stepped away for a few months and then came back, it looked like you needed to learn 15 different systems just to be able to use Claude code effectively. So Claude finished its analysis and it came up with a few clarifying questions to help me think through the scope of what should go into this feature and maybe what should be left out of it. Looks like to start off it has three questions for me to to think through and give some answers on. And by the way, this clarifying questions, this Q and A interface, this wasn't in Cloud Code originally earlier in the year, last year. And so that was a key part of what Agent OS provided since day one. A way to have the agent interview you to really shape the spec before it locks in the plan and before it starts building. But now as you can see that Q and A interface and the ability to think through and prepare really helpful constructive
05:52clarifying questions, All of that now is a first class feature of both Claude code and the Opus 4.5 model. So I'm going read through these questions and give them some thought and I'll give some answers. It's even suggesting a couple different charting libraries and it recommends one over the others. So now based on my answers and based on its continued analysis of my code base and our product, it's going to formulate a plan and then present it to me to review. So I'm going let it cook on that for a minute. And by the way, this is spec driven development. It's now a first class workflow built right into Cloud Code. Something that, you know, I really think is essential for any professional builders in back in 2025,
06:32but also here in 2026. And by spec driven development, I mean, you know, really emphasizing the effort and the thought and the craft that goes into the planning phase before we have the agent go off and actually build. It really makes a big difference in the level of quality and success rate that that we can have at the end of the process. Okay. It looks like its plan is ready for me to review. So I'm actually gonna read this in detail and and see if I agree with everything that it's including in the scope and its implementation plan. Alright. So it's mostly good but I do have a couple of notes to try to simplify the scope a little bit before we get going.
07:08So I am gonna scroll down and instead of saying yes, go ahead and build, I'm gonna go to the third option here and type some feedback to have Claude update the plan. On the gem insights, let's just focus on tracking number of gems per stream. When it comes to filtering, let's keep it simple and just include the ability to filter by stream and So, you know, I'm working my experience into this knowing that like there are some aspects of this that could be a little bit more complex.
07:38HOOKI'm sure Clog Opus 4.5 could probably handle it just fine, but I I do want to keep the feature somewhat simple, no need to bloat the code base unnecessarily. So Claude is gonna go ahead and simplify the the plan with my feedback. Okay. So the plan has been updated pretty quickly and yeah, it all looks good. So I'm gonna go ahead and give Claude the go ahead to bypass asking me for permissions for every little thing and it's gonna go ahead and start building. So Claude's getting to work and it created a to do list for itself. Again, that's something that Agent OS used to do as well. It would break a spec out into a task list, but now that's just sort of built into the way that Claude code works by taking a plan or a spec and breaking it out into tasks that it can track over a long period of time. So we're gonna let this work for a while and then we'll come back to it. So what changed here in 2026?
08:26HOOKWell, two things converged. First, the models got dramatically better. Opus 4.5 specifically, but also Gemini three and GPT five. Now that we're a few months into working with these frontier models, it's clear now that we're truly at a turning point both in terms of what the models are capable of and an observable increase in adoption. The models understand our intent better and they maintain context more reliably across really long sessions and they execute complex multi file changes with significantly fewer mistakes. The back and forth cycles where you're constantly fixing and re prompting and redoing things, those have dropped way down. Not to zero but enough that the workflow feels fundamentally
09:07HOOKdifferent now. Second, Clogcode itself evolved. Plan mode overall is polished and reliable now for powering a spec driven development approach. And features like skills and sub agents give you targeted capabilities without the framework overhead. So the bottom line is much of what we built scaffolding for back in 2025 is now just part of Claude code. Okay. So Claude is all finished with its work. So let's let's see how we did. Alright. So I see it added the new trends view here in the navigation. Yeah. I see a functional trends view with with some graphs but there are some front end UI issues that I think we're gonna need to smooth out. So I will need to give it a bit of feedback to fix some of these. And I think part of the reason why I'm seeing this is because I initially forgot to invoke the front end design skill which I have in my system but forgot to include that in the in the plan and then I inserted it into Claude's work midway through.
10:04So you can see here, I said be sure to use the front end design skill when designing new UI interfaces for this feature. So that was my mistake. I should have done that earlier before Claude even began its work. And so I think that like inserting that midway through might have thrown it off a little bit and resulted in a little bit of inconsistency here. But this looks like the kinds of things that I think we can probably smooth out with a bit of feedback.
10:31Alright. So I pointed out a few UI issues and I provided several screenshots along with those. So, let's have Claude work on those fixes.
10:43Alright. So, it took a few minutes to fix those UI issues. Let's restart the server and see how we're doing. Okay. Yeah. This is much better. Yeah. It's totally in line and consistent with the other pages. So those layout issues are fixed and it's now using the consistent card style that we've already established in the app and let's see how we're doing on mobile. Looks great
11:04and everything in the app is dark mode including this trends view. So this is perfect. I would say this is pretty close to a shippable state. Now, me show you a couple of core features in Claude code that really matter when you're doing professional work like this. Claude skills are becoming an essential tool. Now, you can still do a lot without them, but when you need specialized workflows or capabilities,
11:27Claude skills really help you up your game. For example, I just invoked the front end design skill to build this trends feature. The Claude team actually created that one and it adds some design muscle to the front end design process. Another one is context management. You can always pull up older sessions to pick up where you left off or rewind if your context gets off track. Small power features like this in Claude code really make a big difference in productivity. Now, I want to be clear about something. When I say that Claude code handles 90% of the work, I don't mean that I just type a prompt and magic happens. Our experience as builders matters more than ever. It's just applied differently. The models can implement any pattern we describe but they can't choose which pattern is right. They can't understand your users. They can't make strategic calls about what to build and why. Product thinking and knowing how to formulate real solutions for real people. Understanding the difference between what a customer says that they want and what they actually need. That's the craft now and honestly,
12:27that skill will never be obsolete. It's something that we should always be working to improve. There's always another level, especially now. The bottleneck isn't writing code anymore. It's knowing what to build and how to structure it. And that's where our . And so when I say that Claude code is all you need, I mean that the tooling is all you need. But your judgment, your taste, your product instincts,
12:49that's what makes it work. Now, said that Claude code and nothing else could handle 90% of the work. But what about the other 10%? Well, are still some scenarios where some additional structure could help. One is greenfield product design. So when you're starting from zero and you need to establish a design system and UI patterns and an architecture before you have a code base to work in. And that's actually why I created design OS. Yeah. And sure that's a free framework y thing but really it just gives you a structured process for using Claude code in the design phase before the actual building begins. Another scenario is when you're working with legacy code bases where you have established conventions
13:27and having those standards documented and then easily referenced by agents can help bring those projects into today's AI first workflow. And that's an area where I think Agent OS could actually help. But here's the important takeaway. You should notice when you need something more than just clog code because you'll feel that friction. The mistake is assuming that you need extra tooling before you allow yourself the opportunity to be productive using just Cloud Code and nothing else. So whether you're just starting to use Cloud Code or you've been deep in frameworks and custom setups, the path in twenty twenty six is the same. Start with pure Cloud Code, learn the native features, but more importantly, your workflow.
14:07CTAOnly add complexity when you actually need it. Now, mentioned Agent OS and I built it to solve real problems that we felt back in 2025. But looking at it now in this landscape, I'm rethinking what if anything I actually want out of a utility framework like Agent OS. And so if you've been layering tools and frameworks on top of your Claude code workflow, then I want you to see my next video. I'm going to show you exactly how I'm stripping away the unnecessary bloat and getting back to what actually matters. It's about subtraction not addition
14:39CTAwithout losing touch with what makes our craft of building software unique. So if you haven't yet, hit subscribe on the channel so you don't miss my next video when it comes out and I'll see you over there next. Let's keep building in 2026.
— full transcript
§ 05 · For Joe
Steal the subtraction frame.
Builder positioning playbook
The creator of the leading Claude Code framework is telling his audience to stop using frameworks — and it works because it leads with the outcome builders actually want.
- Open a video by disqualifying your own product — it builds instant trust and filters for serious builders.
- Use the 90/10 rule as a positioning anchor: 90% native, 10% intentional structure.
- Frame spec-driven development as the default professional workflow: plan, clarify, review, build.
- Invoke design skills before the plan, not during. Brian learned this on camera; worth sharing as a tip.
- The craft argument is your content angle: AI handles implementation, you supply product judgment.
- For JoeFlow/MCN+ positioning: the tool handles the session, you handle the intent. Same frame, different product.
§ 05 · For You
You do not need to master the ecosystem first.
If you are just getting started
The most expensive mistake in AI-assisted development is spending months learning frameworks before you ever ship anything.
- Start with vanilla Claude Code: plan mode, voice your questions, let it ask clarifying questions before it builds.
- Only add a framework when you feel the friction of needing it — not before.
- Product thinking matters more than prompt engineering. What to build and why is still your job.
- Spec-driven development is not a framework. It is just making a plan before you start.
