Latest / AI Fire Daily / #533 Neil: Fable 5 Builds A Cinematic Website Inside Claude Code
Transcript
- 0:00Imagine just speaking a luxury website into existence.
- 0:05Beat. You don't write a single line of CSS. You
- 0:07don't wrestle with JavaScript. You just describe
- 0:09a cinematic vision, and you watch it materialize
- 0:12on your screen. It feels almost like magic, you
- 0:14know? But the reality of how it happens is actually
- 0:17far more interesting than magic. Oh, absolutely.
- 0:20I mean, it completely upends the traditional
- 0:22web development model. We are moving from manually
- 0:26stacking bricks to Essentially architecting the
- 0:29entire building through specialized language.
- 0:31Yeah. And the technical mechanics of how that
- 0:33translation happens from human intent to a fully
- 0:36functioning DOM structure is just fascinating.
- 0:40Welcome to the Deep Dive. We are unpacking a
- 0:42rather brilliant source guide today. It explores
- 0:45how to use Fable 5, specifically inside Claude
- 0:48Code, to build a high -end cinematic website.
- 0:51That's a really good read. Yeah. The journey
- 0:53we are taking today starts with structuring your
- 0:55initial vision. From there, we are going to break
- 0:57down the architecture of the megaprompt, examine
- 0:59the AI's iterative build process, and finally,
- 1:02look at the reality of deploying what you've
- 1:04made. And we should probably note the timing
- 1:06here. Anthropic has extended Fable 5 access on
- 1:09paid plans. Right. They're giving 50 % higher
- 1:12weekly rate limits through July 19, 2026. Wow.
- 1:18Yeah, so this isn't just an abstract theory anymore.
- 1:21It is a highly accessible window for anyone wanting
- 1:24to test the absolute edges of AI front -end generation.
- 1:28Which means this is the perfect time to experiment.
- 1:31Let's start with the foundation. There is a crucial
- 1:33realization you have to make before you even
- 1:35open a terminal window or a chat interface. AI
- 1:39does not read your mind. It definitely doesn't.
- 1:41No. It demands meticulously curated inputs. You
- 1:45are essentially setting up the cognitive boundaries
- 1:47for the model. That is the perfect way to frame
- 1:49it. The guide actually grounds this in a fictional
- 1:52test project called Aurum Noir. Right. The watch
- 1:55brand. Yeah. It's a Swiss luxury watch brand.
- 1:57And the point they make is you cannot just prompt
- 2:00the model to build a Cool Watch website. No,
- 2:03you can't. You are going to get boilerplate garbage.
- 2:05You have to lock down three strict prerequisites
- 2:08before you even begin. Let's walk through those.
- 2:10The first one is a crystal clear concept. Right.
- 2:13You need to know the target audience, the exact
- 2:15feeling you want to evoke. For Oram Noir, it's
- 2:19premium and minimal. Yeah. And you need a primary
- 2:21call to action. Here, it is book a private viewing.
- 2:24The second prerequisite involves visual references.
- 2:28You need two or three inspiration sites saved
- 2:30in a folder. We are talking about highly specific
- 2:34layouts or typography styles you want the AI
- 2:36to emulate. The model needs a baseline of what
- 2:39you actually consider good. And the third piece,
- 2:42which I find the most interesting from a systems
- 2:43perspective, is organizing your assets. Yeah,
- 2:46this is huge. Your logos, founder portraits,
- 2:48selected fonts, and the brand copy. But the guide
- 2:51emphasizes creating an asset map. And AssetMap
- 2:54is just non -negotiable for a project of this
- 2:57scale. You are essentially giving the LLM a spatial
- 3:00understanding of the file directory. It tells
- 3:03the AI precisely where each image or block of
- 3:06text should be injected into the HTML structure
- 3:08later on. I think about this preparation phase
- 3:11a lot like Miss On Place in a high -end kitchen.
- 3:14Oh, that's a good analogy. You know, you don't
- 3:16just throw raw ingredients at a master chef and
- 3:18hope a Michelin star meal comes out. Exactly.
- 3:21The prep you chop, you measure, you organize,
- 3:23you curate the environment. But I have to make
- 3:27a vulnerable admission here. I still wrestle
- 3:30with prompt drift myself. Oh, everybody does.
- 3:32It is the most common failure point. Yeah. Prompt
- 3:35drift happens because of how context windows
- 3:38function. If you don't anchor the model with
- 3:40a strict asset map and a rigid concept up front...
- 3:44It just wanders off. Right. The AI starts hallucinating
- 3:47intermediate steps. Its attention mechanism wanders.
- 3:50It might, you know, invent a new font halfway
- 3:52down the page because it just lost the thread
- 3:54of your original stylistic constraints. To counter
- 3:57that, the guide suggests a specific technical
- 4:00setting before you begin. You need to set Fable
- 4:025's effort level to high. That is a crucial flag.
- 4:07Setting the effort to high forces the model into
- 4:09a more deliberate reasoning state. It makes the
- 4:12AI thoroughly review your 10 -section structure
- 4:16and your asset map conceptually before it attempts
- 4:19to write a single line of executable code. It
- 4:22is basically allocating more compute to planning.
- 4:26There is another catch here regarding assets,
- 4:28specifically images. Oh yeah. The guide is adamant
- 4:31about avoiding text baked into product images.
- 4:34no labels drawn onto the JPEGs. Right. Now I
- 4:37understand that from a traditional web design
- 4:39standpoint, but why is it so critical when using
- 4:42an AI builder? Because of how the AI constructs
- 4:45the document object model, the DOM. If your text
- 4:48is trapped inside a flattened image file, the
- 4:50AI cannot manipulate it. More importantly, it
- 4:53destroys accessibility. Screen readers are completely
- 4:56blind to a JPEG of a button. Yeah, that makes
- 4:58sense. Plus, it breaks your mobile responsiveness.
- 5:00If a user shrinks their screen, baked -in text
- 5:02shrinks to an illegible smudge. All labels and
- 5:05interactive elements just have to be rendered
- 5:07as real HTML and CSS. I hear that, but let me
- 5:11challenge this entire setup phase for a moment.
- 5:13Okay. If these models are so incredibly advanced,
- 5:16why shouldn't we just let Fable 5 design the
- 5:18layout from scratch? Wouldn't giving it a blank
- 5:21canvas save us a massive amount of preparation
- 5:23time? It feels intuitive to just let the machine
- 5:26do the thinking, right? But a blank canvas prompt
- 5:29forces the AI to regress to the meme. How so?
- 5:33Well... LLMs are trained on vast datasets of
- 5:36the entire internet. The average website is terrible.
- 5:39It's generic. Fair point. If you don't constrain
- 5:42the AI, its probabilistic nature will just output
- 5:45a mathematical average of every standard boilerplate
- 5:48template it has ever seen. The constraints are
- 5:51what force it into the narrow, high -end probability
- 5:54space you actually want. So constraints actually
- 5:56fuel the AI's creativity instead of limiting
- 5:59it. Precisely. You are shifting your role from
- 6:01a front -end developer to a creative director.
- 6:03The boundaries you set are the actual design
- 6:05work. Two secs silence. Let's move to the second
- 6:08phase of this process. We have our myths in place,
- 6:10we have our constraints, now we need the master
- 6:12blueprint. This is where we craft one massive
- 6:15hyper -detailed prompt. Step 3 in the source
- 6:18guide revolves entirely around this mega -plompt.
- 6:21The architecture of this prompt is very deliberate.
- 6:23You provide the entire vision in a single comprehensive
- 6:26injection rather than dripping it out over 20
- 6:30small requests. The granularity required here
- 6:32is intense. The prompt dictates a 10 -section
- 6:35structure. Hero, manifesto, craftsmanship, the
- 6:39movement, the collection, limited edition, heritage,
- 6:42collector's words, private viewing, and the footer.
- 6:45It's a lot. And it also locks in the precise
- 6:48brand styling. You aren't just saying make it
- 6:50dark. You're defining a near -black background.
- 6:53specific hex codes for orange accents, gold detailing,
- 6:57and ivory text. It even pairs the typography.
- 7:00Cormorant Garamond for the headlines to establish
- 7:02that classic, luxurious serif weight, and Manrope
- 7:05for the body text to maintain a clean, highly
- 7:08legible, modern contrast. But the true power
- 7:10of this mega product is how it handles the interaction
- 7:13design. You are explicitly asking for slow, cinematic
- 7:16scroll experience. Yeah. You define the sticky
- 7:19sections, you ask for soft parallax. Let's quickly
- 7:21define that for clarity. Backgrounds moving slower
- 7:24than foregrounds to create a 3D depth illusion.
- 7:27Exactly. It is a subtle effect, but it is the
- 7:30hallmark of luxury web design. The prompt also
- 7:33specifies text reveals as elements enter the
- 7:36viewport. Right. And perhaps the most vital instruction.
- 7:40It tells the AI not to use external media services
- 7:43or heavy video files to achieve this. This brings
- 7:45up a serious question about how we interact with
- 7:47LLMs, though. OK. Looking at this massive wall
- 7:50of text. Wouldn't it be far less prone to error
- 7:53if we built this piecemeal, say, prompt the heroception,
- 7:56verify the code, then prompt the manifesto, verify
- 7:59that, and so on? That is how most people naturally
- 8:02interact with chatbots. But it fails terribly
- 8:05for cinematic web design. Why is that? Think
- 8:07about the cascade of CSS and global state management.
- 8:10To build a cohesive, slow scroll experience where
- 8:13a sticky section elegantly transitions into a
- 8:16parallax manifesto, the AI needs the global context
- 8:19of the entire DOM tree. Right. If you build it
- 8:21sequentially in isolation, the AI loses sight
- 8:24of the overarching aesthetic rhythm. The transitions
- 8:27become jarring and decoupled. That makes total
- 8:29sense. It needs the full architectural blueprint
- 8:32to understand how the plumbing connects between
- 8:34floor 1 and floor 10. Yes, perfectly said. But
- 8:37speaking of plumbing, the prompt asks for a book,
- 8:41a private viewing call to action. How is Fable
- 8:445 supposed to handle a booking form without a
- 8:48real database hooked up to it? This is where
- 8:50you have to explicitly manage the AI scope. The
- 8:53prompt instructs Fable 5 to build front -end
- 8:56validation only. Oh, I see. It tells it to construct
- 8:59the form, validate that the user entered an email
- 9:02address, and display a polished success message.
- 9:05You were instructing it not to hallucinate a
- 9:07fake backend or try to wire up a dummy API. Right.
- 9:10It builds the illusion of a working form, but
- 9:12no backend exists yet. Exactly. It isolates the
- 9:14build strictly to the user interface. It lets
- 9:17you perfect the client -side experience without
- 9:19tangling the AI and server -side logic that you
- 9:21don't even have an environment for yet. Before
- 9:23we move into the actual build phase and see how
- 9:25the AI executes this massive prompt, a quick
- 9:28word from the folks who make this deep dive possible.
- 9:31We talk a lot about execution on this show, and
- 9:34whether you are building a cinematic front end
- 9:36or a massive data pipeline, having a reliable
- 9:39infrastructure is critical. That is why we are
- 9:42thrilled to partner with our current sponsor,
- 9:44who provides the seamless cloud environments
- 9:46necessary for modern development. When your tools
- 9:49just work, you can focus purely on your creative
- 9:52vision. And that transition from vision to execution
- 9:55is exactly what happens next in Cloud Code. We
- 9:58are back to the process. We just fed Fable 5
- 10:01the master blueprint. This next phase is where
- 10:03we transition from theory to execution. And honestly,
- 10:07watching how the AI writes the code and self
- 10:09-corrects is just staggering. Fable 5 takes that
- 10:11mega prompt and begins generating all the HTML,
- 10:14CSS, and JavaScript. And it does it within the
- 10:16context of your local file system, creating the
- 10:18directory structures and linking the assets you
- 10:21mapped out. There is a very specific technical
- 10:23nuance the guide points out here. You do not
- 10:26need to ask the AI to lean on heavy animation
- 10:28libraries. No GSAP, no framer motion, no Linus,
- 10:32or 3JS. People often assume you need those heavy
- 10:35JavaScript libraries for cinematic movement.
- 10:38But modern vanilla CSS is incredibly powerful
- 10:42now. Things like CSS scroll timeline and native
- 10:45intersection observers can handle complex reveals
- 10:48and sticky positioning. Fable 5 understands this.
- 10:52It's wild. Unless you are asking for highly complex
- 10:553D object manipulation, it can achieve that premium
- 10:58feel using native browser capabilities, which
- 11:01keeps the site incredibly fast. It handles the
- 11:03desktop and mobile layout simultaneously, writing
- 11:06the media queries as it goes. But here is the
- 11:08mechanism that genuinely stopped me in my tracks.
- 11:10What's that? While it is building, Fable 5 doesn't
- 11:12just write code blindly. It spins up a live version
- 11:15of the site and reviews its own work. It runs
- 11:1736 automated checks against the live DOM. Whoa.
- 11:21Imagine it running 36 automated QA checks in
- 11:23seconds. It is hard to wrap your head around.
- 11:26How is it fundamentally doing that? Is it just
- 11:29reading its own text output? It is far more advanced
- 11:32than that. It is essentially using a headless
- 11:35browser, a web browser running in the background
- 11:37without a graphical user interface. OK. It renders
- 11:40the code, parses the actual live DOM, and runs
- 11:44assertions against it. So it's testing everything?
- 11:46Yes. It checks the structural hierarchy. It tests
- 11:49the contrast ratios for accessibility. It measures
- 11:52the loading waterfall. It verifies that the call
- 11:55to action buttons are actually clickable. And
- 11:58it isn't just generating a passive report. In
- 12:00the guides test project, Fable 5 found a duplicate
- 12:03font download issue during these checks. That's
- 12:06a great catch. It realized two different CSS
- 12:08rules were fetching the same heavy font file,
- 12:11which causes render blocking delays. It autonomously
- 12:14rewrote its own CSS to consolidate the request,
- 12:17verified the fix, and continued building. That
- 12:20self -reflective loop is the true lead forward
- 12:23here. It is acting as both the junior developer
- 12:25writing the code and the senior engineer conducting
- 12:28the code review simultaneously. But this phase
- 12:31also introduces a major friction point. The guide
- 12:35calls it the sandbox trap. Oh yeah, this trips
- 12:37up almost every first time user. When Cloud Code
- 12:40finishes the build and runs the preview server,
- 12:42it provides a local host link in the terminal.
- 12:45Usually something like 127 .0 .0 .1 on port 8080.
- 12:49And when you click that link, it just times out.
- 12:52Your browser says the site can't be reached.
- 12:55Why won't that local link just open on the user's
- 12:57computer if the code is being generated right
- 12:59there? It comes down to how Anthropic handles
- 13:02security and environment isolation. Cloud Code
- 13:05isn't just executing random scripts natively
- 13:07on your laptop's operating system that would
- 13:09be a massive security vulnerability. Instead,
- 13:11it spins up a secure remote cloud container,
- 13:14a virtual machine. The website preview is running
- 13:17inside that isolated cloud environment. So when
- 13:19it gives you a 127 .0 .0 .1 link, it means local
- 13:24to the AI's virtual machine, not your physical
- 13:26laptop. Got it. The local link is trapped inside
- 13:29Claude's own virtual computer. It is an air -gapped
- 13:32environment. To actually view the site, you have
- 13:34to bridge that gap. You have to explicitly prompt
- 13:38Fable 5 to publish a private artifact review.
- 13:41The fix is surprisingly simple once you know
- 13:43the architecture. You just tell it. Publish the
- 13:45current OrmNOR website as a private artifact
- 13:48preview. Give me one direct link that I can open
- 13:51in my browser. Exactly. That command packages
- 13:54the current state of the virtual container and
- 13:56exposes it via a secure external URL. And then
- 14:00it works. You click that artifact link and suddenly
- 14:03you can experience the site natively in Chrome
- 14:05or Safari. You can actually feel the slow cinematic
- 14:09scroll and watch the text reveals happen. Beat.
- 14:12Which brings us to the final phase. Step six.
- 14:15You are looking at your private preview. It looks
- 14:17stunning. But we need a reality check on what
- 14:19it actually takes to finish a project like this.
- 14:22The initial generation gets you 90 % of the way
- 14:24there, but the last 10 % is where the polish
- 14:26happens. You have to send a final cleanup prompt.
- 14:28This final pass is all about auditing the details.
- 14:32You instruct the AI to sweep the entire code
- 14:34base and replace any lingering placeholder text
- 14:37with your approved assets. You mandate a check
- 14:40on every image size and aspect ratio to ensure
- 14:43nothing is stretched, and you force it to compress
- 14:46any large files that might slow down the initial
- 14:48load time. You also need to explicitly prompt
- 14:51it to verify reduced motion support. Let's dig
- 14:53into that because it is an accessibility standard
- 14:56that often gets ignored. systems have a setting
- 14:58where users can prefer reduced motion. If someone
- 15:01suffers from vestibular disorders, heavy parallax
- 15:04and rapid text reveals can actually cause physical
- 15:06nausea. Wow, yeah! A high -end website must respect
- 15:10that system preference. By prompting Fable 5
- 15:13to verify this, the AI injects prefers reduced
- 15:16-motion media queries into the CSS, which gracefully
- 15:19degrades the complex animations into simple fades
- 15:22for those users. The prompt also demands a rigorous
- 15:25final check on mobile responsiveness. The layout
- 15:28has to remain pristine on a 5 -inch screen. But
- 15:30even after all of this polish, there is a hard
- 15:33truth about what AI front -end builders can and
- 15:35cannot do. Yes. That beautiful artifact link.
- 15:39is fundamentally just a temporary private preview
- 15:43server. It's not a live product. I think about
- 15:45this like building a highly detailed million
- 15:48-dollar concept car. It is sitting there on the
- 15:51showroom floor under the lights. The paint is
- 15:53flawless. The leather stitching is perfect. The
- 15:55aerodynamics are mathematically precise. It is
- 15:58breathtaking. But you still need to drop an engine
- 16:01into the chassis before you can actually drive
- 16:03it on the highway. That analogy hits the nail
- 16:05on the head. The AI constructed a perfect chassis
- 16:08and interior. But public hosting, getting a real
- 16:12domain name, setting up CDN delivery, and building
- 16:14a functional backend database, well, that is
- 16:17the engine. If you want real luxury watch buyers
- 16:20to interact with Arm Noir, you have to take the
- 16:22code out of Claude, host it on a public server,
- 16:24and connect that private viewing form to an actual
- 16:26database that processes the requests. The AI
- 16:29handles the interface, but the architectural
- 16:31plumbing is still your responsibility. At this
- 16:34final publishing stage, when users are migrating
- 16:36the code and doing their final checks, what is
- 16:38the most common blind spot? Where do people fail?
- 16:40It is almost entirely psychological. Users get
- 16:44completely mesmerized by the smooth desktop animations.
- 16:47Sure. They sit there scrolling up and down, watching
- 16:50the parallax effects on a large monitor, and
- 16:52they completely forget to open the developer
- 16:54console to check for hidden JavaScript errors.
- 16:57Or worse, they never test it on an actual physical
- 17:01smartphone. Basically, don't let a beautiful
- 17:03desktop site hide a broken mobile experience.
- 17:06The vast majority of your consumer traffic is
- 17:08going to hit that site on a phone. If the manifesto
- 17:11text reveals perfectly on a laptop but the mobile
- 17:13navigation menu overlaps the logo and breaks
- 17:16on an iPhone, the illusion of luxury shatters
- 17:18instantly. The brand trust evaporates to sex
- 17:21silence. Let's zoom out and look at the broader
- 17:24implications of what we've discussed today We
- 17:26just traced a workflow where an AI built a highly
- 17:28stylized 10 section interactive cinematic website
- 17:31Relying only on a folder of organized assets
- 17:34and a few highly structured prompts It forces
- 17:37us to completely reevaluate the bottlenecks in
- 17:39digital creation It does, but the overarching
- 17:42theme that keeps surfacing for me is that AI
- 17:44tools like Fable 5 are, at their core, incredibly
- 17:48capable execution engines. They can calculate
- 17:51animation keyframes flawlessly, they can write
- 17:54perfect syntax, but the human remains the essential
- 17:57director of the entire production. The model
- 18:00possesses no inherent taste. It has no lived
- 18:03human experience of what the concept of luxury
- 18:05actually feels like to a person. It relies entirely
- 18:08on your curation. It was your choice to demand
- 18:11a near -black background rather than a stark
- 18:13white one. It was your selection of Cormorant
- 18:16Garamond to evoke a specific historical era of
- 18:19craftsmanship. Your logic, your asset mapping,
- 18:22and your structured constraints are the only
- 18:25thing separating a generic, forgettable webpage
- 18:27from a deeply immersive cinematic experience.
- 18:30You might not be manually typing the CSS anymore,
- 18:32but you are entirely responsible for the soul
- 18:35of the project. The technical barrier to entry
- 18:37for building complex, beautiful digital spaces
- 18:39has effectively dropped to zero. But because
- 18:42execution is now commoditized, the premium on
- 18:45human taste, curation, and editorial judgment
- 18:47has never been higher. Which leaves us with a
- 18:49provocative new reality to consider. We are currently
- 18:53using these tools to build static, one -size
- 18:55-fits -all experiences. But if an AI can instantly
- 18:59write all the complex code for a cinematic front
- 19:01end based on a prompt, what happens next? The
- 19:05cycle time continues to shrink. Imagine a near
- 19:07future where the AI isn't just building the site
- 19:10once. What happens when the AI can instantly
- 19:13scrape a competitor's site, analyze a specific
- 19:16visitor's psychological conversion trigger, and
- 19:18dynamically rewrite and rebuild your website's
- 19:21entire cinematic front end in real time for every
- 19:24individual person who clicks your link. Now that
- 19:26is a paradigm shift. The website stops being
- 19:28a destination and becomes a fluid, personalized
- 19:31conversation. Something to think about the next
- 19:33time you sit down to curate your assets. You
- 19:35are no longer just building a page. You are learning
- 19:38to direct a machine that will eventually build
- 19:40reality on the fly. Thank you for joining us
- 19:42on this deep dive.