Big Idea

The argument in one line.

The real test of any new AI model isn't what it can one-shot in a tweet — it's whether it holds up on actual production code, and Claude Fable's answer reveals that planning has become the rate-limiting skill, not prompting.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code on real apps and want a grounded first-build report on Claude Fable beyond the Twitter hype.
You're a solo builder who evaluates new models by shipping something, not by running toy prompts.
You're curious how spec-driven development (scoping doc -> PRD -> milestones) changes what you hand off to an AI agent.
You use or are considering Superconductor for managing multiple Claude Code worktrees in parallel.

SKIP IF…

You're looking for a benchmark comparison — this is one builder's qualitative evaluation on one real task.
You don't use Claude Code or agentic coding tools at all.

TL;DR

The full version, fast.

When Claude Fable dropped, the internet one-shotted toys. This video does something more useful: it runs Fable on a real production feature (external trend tracking for a Rails app) using a full spec-driven workflow — scoping doc in claude.ai, milestone-based PRD, Superconductor worktree, then handoff. Fable builds it in about 41 minutes, 44 tool uses, 58k tokens. Two observations emerge: as models get stronger, the planning phase matters more (Shape becomes the premium skill), and the Refine loop that used to take days is compressing toward zero. The honest catch is cost.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 01:02

01 · Claude Fable first impressions

Contrarian premise: skip the toy demo, run a real build. Shows the Anthropic announcement page for Claude Fable 5 and Mythos 5.

01:02 – 02:52

02 · How I evaluate a new model class

Personal evaluation framework: only test on production-relevant tasks you'd actually ship. No toy prompts as signal.

02:52 – 09:20

03 · Planning the build

Full planning session in claude.ai: scoping doc for Resonance Radar external expansion (RSS generalization across YouTube, LinkedIn, Reddit, X), PRD Creator skill, milestone breakdown before any code is written.

09:20 – 15:57

04 · Handing it to Fable

Opens Superconductor, creates a worktree (resonance-radar-external), hands off the scoping doc + PRD to Claude Code. Fable runs: clarifying questions, ~1:40 exploration phase, 58k tokens consumed.

15:57 – 21:34

05 · Reviewing Fable's work

Live Resonance Radar showing external trending topics with magnitude/velocity/outlier scores. 'My CLAUDE.md after 6 months of daily use' is a 3.1x magnitude trending topic. Final build report: 44 tool uses, 58k tokens, ~41 min.

21:34 – 23:40

06 · 2 early observations

Shape->Build->Refine diagram and 'The model divergence' title card. Two takeaways: planning is the new bottleneck, refinement is collapsing. Honest catch: cost.

Atomic Insights

Lines worth screenshotting.

Toy demos don't tell you if a new model is worth reaching for — only a real build does.
When the model can execute reliably, planning becomes the rate-limiting variable, not prompting.
The refinement stage — once the grind — is compressing toward zero as models get stronger.
The Shape->Build->Refine loop doesn't shrink evenly: Shape is expanding in importance as Build collapses in effort.
A scoping document created before handing off to Claude Code is more valuable than any prompting technique.
Model divergence is real: different models have meaningfully different execution profiles on the same spec.
Evaluating a model on a feature your business actually needs is the only reliable signal.
44 tool uses and 41 minutes is a benchmark worth knowing for a mid-sized Rails feature with external API integration.
The bottleneck shift is a skill shift — builders who plan well will outperform builders who prompt well.
Spec-driven development (scoping doc -> PRD -> milestones) is what makes autonomous AI execution reliable at the feature level.

Takeaway

Planning is the new bottleneck in AI-assisted builds.

WHAT TO LEARN

When the model can execute reliably, the quality of your spec determines the quality of your output -- not how well you prompt.

01Claude Fable first impressions

The meaningful test of a new model isn't what it does in a tweet-sized demo -- it's whether it holds up on a task with real constraints and integration points.
Positioning yourself against the toy-demo crowd signals rigor: your evaluation will be trusted more because the bar was harder.

02How I evaluate a new model class

A personal evaluation framework beats benchmarks: pick a real task from your backlog, run it, and measure output quality against your actual definition of done.
Only production-relevant work surfaces what a model actually can't do -- edge cases, real APIs, existing codebase constraints.

03Planning the build

A one-shot scoping document in claude.ai before opening Claude Code is a forcing function: it surfaces ambiguities before they become wrong code.
Milestone-based PRDs give the model checkpoints to verify against, which reduces the chance of confident-but-wrong execution across long agentic runs.
Planning in a chat interface (claude.ai) and executing in an agentic interface (Claude Code) is a deliberate separation -- the planning step is exploratory, the execution step is constrained.

04Handing it to Fable

Worktrees isolate the risk of agentic execution -- the main codebase stays clean while the model works in a branch that can be discarded or merged.
58k tokens and ~1:40 exploration before first output is a useful calibration: frontier models front-load their thinking, which is good, but budget for it.

05Reviewing Fable's work

44 tool uses in ~41 minutes for a mid-sized Rails feature with external API integration is a useful benchmark for scoping agentic build sessions.
When a model surfaces a tricky integration (Reddit's public JSON vs official API), treating that as a decision point rather than a failure is the right response -- the model found the real constraint.
The result showing your own content as a trending topic (3.1x magnitude) is a live demo that doubles as a trust signal -- the feature works on real data, not synthetic examples.

062 early observations

Shape->Build->Refine is the right frame: as Build gets more reliable, the value of a tight Shape phase multiplies because there's less in Refine to catch your planning mistakes.
Model divergence means the skill of model selection is now as important as the skill of prompting -- knowing which model handles which task class is non-obvious and worth building deliberately.
Cost is not an afterthought: the honest catch on frontier models is spend, and any evaluation that ignores cost is incomplete.

Glossary

Terms worth knowing.

Resonance Radar: A custom Rails app built by the creator for tracking content idea resonance — essentially a personal trend-detection and idea-curation tool with kanban-style organization.
Superconductor: A third-party macOS app for managing multiple Claude Code worktrees in parallel, making it easier to run simultaneous agentic build sessions without context collision.
Shape -> Build -> Refine: A spec-driven development framework where Shape is the planning/scoping phase, Build is autonomous AI execution, and Refine is the iteration loop after first output.
PRD Creator: A free Claude Code skill (buildermethods.com/tools) that generates a milestone-based Product Requirements Document from a one-sentence project scope, used as the handoff artifact for Claude Code execution.
Scoping document: A one-shot artifact generated in claude.ai that captures the full feature spec before execution — what's in scope, what's out, the tech constraints, and the definition of done.
Model divergence: The growing gap in capability and behavior between frontier AI models, making model selection itself a skill rather than a commodity choice.
Claude Fable 5: Anthropic's most capable widely released model as of mid-2026, succeeding the Opus line, with 1M context window and significantly stronger autonomous execution than prior generations.
Worktree: In git and Claude Code, a separate working directory that shares a repository but runs independently — used here to isolate the feature branch from the main codebase during agentic execution.

Resources

Things they pointed at.

NaN:NaNBuilder Methods Pro ↗

NaN:NaNBuilder Briefing (free newsletter) ↗

NaN:NaNBuilder Methods Tools ↗

NaN:NaNPRD Creator skill ↗

NaN:NaNSuperconductor (Claude Code worktree manager)

NaN:NaNRSS.app ↗

Quotables

Lines you could clip.

00:05

“Everyone's one-shotting toys with it. Instead, I put it to work on a real tool my business actually needs.”

↗ Tweet quote

21:40

“Planning matters more than ever -- because the model can execute. The bottleneck shifted upstream.”

↗ Tweet quote

22:20

“The refinement stage is shrinking. What used to be the grind is compressing.”

↗ Tweet quote

23:00

“The honest catch: the cost.”

↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

00:00Build me an app. You can type that into the new Claude Fable model and you can be pretty impressed off the bat. Now, Anthropic just dropped this new class of model.

00:09It's actually the Claude Mythos model, but they're calling it Claude Fable and apparently, it's safer, I think. So my feed this week is full of people one shotting video games and making perfect clones of Slack and Notion with a single prompt using Claude Fable.

00:26And yeah, it's genuinely impressive. I mean, the benchmarks are calling it a real step change above models like Opus. But, know, I've never been able to judge a new model just by toy demos on X or YouTube.

00:39The real test is what it does for our actual work. So today, I'm putting Fable to work on a real tool that my business actually needs, and you'll get to ride along. So by the end of this video, I'll help you figure out how to judge this new class of AI models that are just coming out now.

00:55And also, we'll break down when this Claude Fable model is even worth reaching for because this thing is expensive. So here's the real gap that I'm trying to fill with what I'm building here today. You know, I wanna be a better teacher on this channel.

01:08But currently, the only way that I'm able to keep a pulse on which topics actually matter to you is just from me, like, scrolling Twitter and scrolling my YouTube feed. So I miss some really good topics that I could teach and cover here. What I wanna build is a kind of radar system, like something that can watch the feeds and the sources that I care about and surface topics that are climbing in popularity.

01:31Now it's a good thing that I'm building this today because as of June 22, so less than two weeks from now, Claude Fable will no longer be available under the Claude Max subscription plans. So after that, you're paying for API tokens. And Fable is hungry, about twice as much what Opus costs.

01:49So if you're on the Max plan and you have something real that you wanna build with Fable, right now is the time to try it out. You know, when I show you my real projects that I'm building in my business like I'm doing today, what I really want you to take away from this is to help you start to look at your own business differently.

02:05You know, look for the gaps, look for the busy work or the tools that you wish existed or were too expensive to custom build before. Because building your own tools and agents is now within reach. And I don't mean just for developers, I mean anyone.

02:19You don't have to be technical. Now, of course, you can't just say build me an app and expect it to get it right. I mean, Clock Fable will impress you.

02:26It might even shock you in what it's actually capable of. But that doesn't mean that it will inherently know what you need or what your business actually needs out of your tools. So you or me as the human in the loop, our role in this process is in that early planning and shaping phase and that is more important than ever.

02:44So as always, that's the approach that I'm taking with the tool that I'm building today. Now with Cloud Fable as the workhorse. Let's go.

02:52Now, actually, what I'm building today isn't an entirely new app. I've already built an app called Resonance Radar and I've been using this to organize all of the new ideas that me and my AI agents pitched to me for me to kind of review and green light them into development, into topics that I might share here on YouTube or in my newsletter.

03:13So what we're gonna be building here today with the help of Claude Fable is a major expansion of this app that I use for curating and planning out new ideas. So, as you can see, I already have it set up to sort of extract potential content ideas from a variety of sources.

03:33And this is kind of the UI of where I manage sources. But as you can see, the sources that I currently pull from or my agents pull from when they're pitching ideas to me come from like my interviews with customers, logs from my daily work. It takes a look at the content that I already published, like in my newsletter, my tweets, my YouTube videos, and other survey responses from my audience.

03:56So all of this stuff is very internal facing. So now we need to build out an external engine for monitoring and keeping our pulse on what's happening in this space of building with AI.

04:12So my starting point for any major expansion of an app or when I'm planning out a new app is to actually go into claud.ai. And this is where I spend a lot of time doing very high level exploration and planning and trying out ideas and seeing how I think how I feel about them and thinking through them and then backtracking and doing a lot of back and forth before I have a really firm scope and spec.

04:40I do it all here in claw.ai. So earlier today, what I did was I started with a pretty long sort of brain dump. This is just a raw getting my ideas out.

04:49I spoke into the editor here. Just some like specific requirements for exactly how I'm gonna need a tool like this to work in my workflow and make it work for my business.

05:01So I have some ideas for this going in. And then I went through a long back and forth with Claude. So this took me, you know, more than half a day to really go back and forth.

05:12And as you can see, what we're doing is we're not building the actual plan yet. We're making decisions together. So this is me using Claude as a thought partner, as a strategic planner, as I think through a major tool expansion and something that I'm building here.

05:29So we looked at different things like what's going to go into the UI and the UX, but also preliminary technical research into what's actually possible for us to pull information from the YouTube API. I wanna pull information off of, like, x, you know, Twitter feeds, and I was exploring LinkedIn and Reddit to see if I can pull ideas from those sorts of feeds to understand what are the trending topics in these platforms, What are the top questions that people in the audience have?

06:01And these are things that potentially I could talk about and teach here on the channel. So again, we went through a long back and forth just hashing it out. And for most of this, I was actually using the new Claude Fable model and that is available here in the Claude model picker.

06:20But as you can see, there's this caveat that it's included until June 22. So as of today, I'm recording this on June 10. That's like less than two weeks away.

06:29So I better get my Fable usage in now before it runs out. So we kept going. As you can see, this is a really long one.

06:35They tend to be pretty long when we're planning out a new product. My output for this session is this artifact.

06:41So I'm always kind of building toward a usable artifact. This is not quite what I would consider to be a PRD.

06:50This is more like a really extensive initial prompt that I'm going to copy and paste into Claude code in just a minute.

06:59So this is kind of like locking in all of the decisions that me and Claude made together in this in this planning session. We sort of locked them in and codify them into this scoping document. I specifically asked Claude to try to avoid being too prescriptive, like how to code this up, how to actually implement it.

07:20This is more about defining the scope, what we are going to be building here, what we're not building, what we're going to be integrating with, and a couple of other like important points here. There is a little bit of database modeling here.

07:34Again, it's pretty high level. So like establishing some new entities that we're gonna be adding into my application. Now, one of the most important pieces in this in this document here is this final section, verification criteria.

07:48Definition of done. So this is becoming a really important pattern in how we spec and plan and give AI models, especially the really powerful ones like Claude Fable.

08:03How do we get give them these initial instructions or our initial request and give them a set of criteria, like literally a checklist here that it can use to define done? Because I'm just gonna let Claude Fable, you know, cook on this and it's gonna go.

08:21And instead of my usual thing, which is to break it up into, you know, smaller buildable chunks, I'm gonna see if it can actually build all of this. And as it goes, it can use this as its verification to really self check and self test. And that's when it knows it'll be done, and we'll see how far we go with it.

08:38So now I have something that I can actually copy and paste into Claude code. And that's what we're gonna do here in just a sec. Hey, real quick.

08:47If you're new here, I'm Brian Castle. I help builders stay ahead of the curve as building with AI becomes the norm. If this is the kind of thing that you wanna build yourself, I teach a course called become the builder.

08:58It's the missing on ramp for building real apps with AI, even if you've never coded before. So we cover the foundations, the tech stack, the templates, and you'll build and ship your first real app.

09:09That's inside of BuilderMethods Pro. That's where you'll get access to all my courses along with our community of builders.

09:16So you can join us by going to buildermethods.com/pro. I hope to see you in there. Okay.

09:20So we have this shaped up scoping document from my conversation with Claude. I'm going to copy that and then I'm gonna hop into Claude code. So here we are in Claude code.

09:31As you can see, I am running the Fable five model currently with high effort. And once again, we have the this little announcement from Anthropic that Fable model is here and it's included in my plan, my Claude Max subscription plan until June 22.

09:48So I better start building now. Now, I'm currently on the main branch, so so I'm actually going to spin up a new work tree for this. So now I'm actually operating in a new work tree.

09:58By the way, I've been jumping around different apps for managing agents in different work trees. Today, I'm using one called Superconductor, which I really like so far.

10:06It's a really nice interface, especially if you're doing, you know, multiple work trees, multiple agents across multiple projects like I tend to do. Anyway, so here we are.

10:15Now, as an experiment, I wanna see if I could sort of skip ahead and skip a few steps that I would normally otherwise do in my building workflow. So normally, I would take this shaped up document and I would bring it into Claude Code, but from there, I would then turn it into more of a formal PRD or product requirements document, which is sort of like a buildable road map.

10:43And actually, I have a free agent skill that you can plug into Cloud Code or any agent called PRD Creator. And I covered that in-depth in a video on the channel just a couple weeks ago. That skill is available at theirmethods.com/prdcreator.

10:58So this is what I normally use these days to plan out a new build. I would take my scoping document or a raw initial prompt and it would turn it into a p r d and then actually break it out into multiple milestones. And then I would feed each milestone into Claude code.

11:15And I would also run plan mode on Claude code to really have it plan out the implementation and kind of go one piece at a time.

11:24And that's how I tend to like to build. I probably still will continue to build that way. But today, I want to be a little bit more ambitious and just really put Claude Fable to the test.

11:34So I'm just gonna go ahead and paste that shaping document in as is. And I'm just gonna add one more thing, and I'm gonna say here, that's what we're building today.

11:44I want you to ask me any clarifying questions about the implementation before you get started. So let's see how Fable does with asking me clarifying questions, which I think is always a really important step in the process. Now, by the way, again, normally I would probably use like plan mode for something like this.

12:02I might even use the new goal feature in Claude code since this is gonna be a pretty big project for it to bite off. But I kinda wanna just see how it works just, you know, plain old Claude code, drop a huge prompt in there, see how we go.

12:17Now, as you can see, this is not, you know, just a thin build me an app prompt. Uh, you know, I really spent a lot of time in Cloud Code like hashing out exactly what I wanna build. So there's a lot in here, a lot of detail and it's gonna analyze what I already have in my Resonance Radar application.

12:37So I expect that it'll have some questions. And this is one of those things that I like to evaluate a new model to see what kind of questions, to see how constructive and thoughtful they are, to see what it comes back with. Because, again, that pattern of, you know, telling the agent to ask you clarifying questions or to interview you or to grill you, that's a really important pattern because it helps shape your thinking, it helps shape your planning, and that's how we kind of transition from just being a vibe coder to being a professional builder.

13:08That's what this is all about. So this is pretty typical these days. I've noticed this with Opus as well.

13:14When it gets started on a big exploration of a code base or really big, you know, prompting request, it will actually spin off multiple sub agents automatically.

13:24And you can you can literally see that down here. So this is like the main agent. And then over here, it's doing like an exploration of the code base.

13:34So normal stuff, it's thinking, it's checking things out. Okay.

13:38So that initial exploration that took about a minute forty seconds, 58,000 and here are a couple of clarifying questions that it's asking me.

13:48And I'm just gonna get these answered for Claude. These are pretty good questions so far. Ah, it caught a little detail that I forgot to provide it earlier.

13:56Alright. So, you know, it's finding little details that I sort of overlooked or kind of have the wrong information in my original shaping document.

14:04So this is really good. It it's actually catching some things that I think Opus probably wouldn't have caught to be honest. It correctly notices like what my infrastructure is in my Rails application, what kind of background job system I have working and it's gonna incorporate that into some of the logic.

14:20So that's good as well. By the way, I already put in place a YouTube API credentials for the agent to be able to access.

14:29And another note, I didn't mention this earlier. I'm following a very similar pattern that I'm using in almost all my apps these days. And I covered this in a previous video called the night shift model, where I build a custom application like Resonance Radar, which is sort of the UI layer and the API layer.

14:49And then I have AI agents, whether it's Claude Code or my Hermes agent or Claude Cowork, be able to run on recurring schedules and that and those are what do the actual thinking and creative work, and then they read and post data and information back to these custom apps. So that's what we're gonna be doing here. And as part of that shaping document, it was which was it was sort of buried in there.

15:14We're instructing Claude to work on both expanding the feature set of Resonance Radar at the application, as well as building out some new agent skills that I'm going to assign to my agents running on a daily recurring schedule.

15:29Okay. So I can see here that Claude is now starting to just go ahead and get started. Right?

15:34Now, again, I didn't do the typical plan mode, which usually kind of shows me the whole plan and then gives me like a a confirmation to yes, go ahead and get started. Since I just started in regular, you know, prompt mode, it's just gonna go ahead and get started and it's creating it has created a to do list for itself.

15:52So, We're gonna just kinda let it cook here and let's see how we go. Okay. So it's all set and let's see how it did.

16:00So it made this final report and, you know, this reads pretty typical from a typical like final report when I'm using Opus. I would say the writing is a bit tighter and easier for me to quickly get a grasp on what was built. So I like that.

16:15I noticed down here it said that it crunched for about forty one minutes. I think that might have included some of the early planning as well. Right.

16:23So let's actually run the server and see where how we're doing. So I'm still like running this locally. This is with like seeded fake data in here.

16:31So I can already see that it added a bunch of stuff to the navigation. So we have a watch list. So these are like new sources.

16:39So we have the original sources over here, and then we also have these like external sources. Later on, I might wanna like merge these into the same area, but more of a design decision that I might make later, like now that I'm looking at it. So yeah, like a bunch of, like, random YouTube channels and Reddit subreddits that we've sort of like preceded in here.

17:00And each like channel or subreddit has some like details, like what kind of platform it is and the feed URL.

17:07Okay. That's looking good. Alright.

17:09So here is the list of like trending topics that it's sort of extracting from these from the sources, from the watch list.

17:17Now, I don't know if these are actually real. Let's see. View original.

17:21This is actually not correct. Like, it's not even the same topic and I'm also not really seeing any other content other than this.

17:30So I don't know if this might be just like a fake seeded piece of content because we're in the local development environment. I'm gonna ask Claude about that.

17:40One other thing that I'm noticing like now that I'm looking at it, so it did build out like the algorithm if you will for how we calculate how much a topic is trending. Right? Like the magnitude of from the source norm.

17:53That's a little bit confusing, but I know what it means. It's like how much this is trending from how off like, what a typical topic within the subreddit would do and like the velocity.

18:03So how much time goes by before it gets to this level of popularity. But the way that we're displaying this stuff is a little bit confusing. So, yeah, those two issues, let's go back and ask the model, you know, what we can do about these things.

18:17So again, this is like one more test of like how it handles feedback and this kind of refinements loop. So let's just do one quick round here. I noticed that the trending topics that are shown seem to link to real posts.

18:32They don't seem to be found on YouTube and Reddit and X. Are these dummy fake data records or are these supposed to link out to real posts that actually live on the Internet?

18:45Then, you know, sometimes I like to like batch my quick feedback like this. I take a screenshot of that.

18:53Also noticing that these stats for magnitude and velocity and the outlier score, I get what they mean, but they're a little bit confusing.

19:03Can we make these simpler and especially have some visual indication for how these stats compare to the norm or like a I don't know, like a visual graph or heat tracker.

19:21I don't know. Something like that. And let's take a look at the reports.

19:24So here is a report. We've got a couple of featured items and we have some content here. So that's pretty good, but I would have expected the padding on this to be a little bit cleaner.

19:34So I guess it didn't really catch that. Let's have it clean that up.

19:39The padding on this report content needs to be cleaned up. The text is currently butting up against the border. Alright.

19:46So I'm gonna let the model churn on that for a minute.

19:57Now, of course, a lot of this stuff could have been a result of kind of giving it overly specific instructions in my original shaping document. That could have been a risk that I ran here.

20:09And also and also, you know, again, I skipped through my typical process, which takes a little bit more of a stepped approach with a proper PRD and breaking it up into milestones. You know, something like this, we probably would have built in, like, two or three different milestones. So that's my normal process when I take a more methodical, like, hands on approach.

20:29The other thing is, like, since we're building into an existing application, an existing code base, that can always be a little bit tricky. But, you know, I wanted to kind of put Fable to the test here.

20:38Let's see how it's doing. Okay. So actually, those were dummy records that I created purely just to verify the UI.

20:44So I think that those actually weren't bugs. It it's really just, you know, those aren't expected to link out to real posts. Yeah.

20:51So that's something that I'll verify like once we get this deployed onto production. Okay. So I worked on that for about seven minutes.

20:58A little slow for kind of like quick turnarounds like this, but, you know, I don't expect this to be a very fast model. Yeah. So let's see how it did.

21:06Yeah. These are actually much better. More visual and the content of how they work here seems to make a little bit more sense.

21:15I might massage it a little bit more once I start to use it day to day. That'll be like the real test for like how useful this stuff is, especially when I'm looking at this data and using it in my real process and my business. But that's definitely an improvement.

21:27And let's see how it did with the report. Yeah. It just cleaned that up.

21:32So, yeah. That's good to go. So look, we're in the earliest days of getting to know Claude Fable.

21:37I mean, clearly, it is a step change in capability, but we'll see what kind of impact it has in the coming months. For now, I do have two observations that we should keep our eye on. So observation one, that refinement stage is starting to melt away.

21:50You know, for the past year, my whole process has been spec driven. Shape a clear plan, build, and then refine. And that refinement stage, you know, rounds of back and forth after that first build, that's always been a real chunk of work.

22:04But with Fable, I think that part is starting to disappear. Not the planning. The planning matters as much as ever, maybe even more.

22:11It's the fixing and the re prompting that no, not like that stuff is reducing since Fable seems to be really good at checking its own work. As long as you give it clear notes on what done looks like, you know, that verification criteria.

22:24So that brings me to observation number two. Choosing the right model is now the new skill Because this capability is expensive and it's about to get even more expensive after June 22 because Claude Fable is not gonna be covered under the Claude Max plan after that time, at least right now. So the question isn't just can the model build this, now it's a question of is this the job that's worth pointing Fable at?

22:49So I think we're gonna start to see a divergence of what we think of as our daily driver models. I mean, for me, that's still Opus.

22:56But now, we also have these heavy hitter models like Claude Fable, where we can use them in select cases for really big important jobs. Now, being able to leverage AI tools and the best in class models like Claude Fable is pretty incredible. But only if you're able to plan like a professional builder.

23:14The good news is that planning skill and the building skill is learnable even if you've never written a line of code. So I made a whole video walking through exactly how I take a raw idea and shape it into a plan that's ready to build With the same attention to detail that an experienced software designer would bring. So you can go watch that next.

23:32But before you do, hit subscribe so you don't miss my next video on building with AI. And then I'll see you over there next. Let's keep building.

The Hook

The bait, then the rug-pull.

Everyone one-shots toys when a new model drops. This video does something harder: it hands Claude Fable a real production feature — external trend tracking for a live Rails app — with a full scoping document, a milestone-based PRD, and a dedicated Superconductor worktree. What comes back in 41 minutes is a working build. And what that build reveals changes the conversation about where the bottleneck actually sits.

Frameworks