Fable 5 Website Factory: The Prompt Is the Product

Nick Saraev's Fable 5 website experiment is useful, but not for the reason the title suggests.

The surface story is dramatic: Fable 5 is back, Nick gives it a broad creative mission, and the model spins up 25 strange, animated, highly visual websites using sub-agents, asset tools, deployment, and review passes. The more useful lesson is quieter: the best prompt was not a design brief for one page. It was an operating system for generating many directions, checking them, hosting them, and packaging the results so other people could reuse the pattern.

That is the part builders should copy. Not the claim that every AI-generated page is suddenly a "$10K website." The business value is still in strategy, copy, accessibility, performance, analytics, deployment, and taste. But Fable 5 makes the exploration phase feel very different.

Video credit: Nick Saraev. The transcript mentions a free /guide route with templates and prompts, but no stable public guide URL was included in the supplied material, so I am linking the video and Nick's X profile rather than guessing.

Source Note

This post uses the supplied transcript from Nick Saraev's video, plus public sources from Anthropic, Claude Code, Chrome DevTools, Higgsfield, and Netlify. I treat the 25-site run, the rough timing, and the per-site token spend as Nick's reported experiment. I treat Anthropic's pages as the source of truth for Fable 5 availability, pricing, safeguards, and Claude Code capabilities.

One timing caveat: Fable 5 access and included usage windows have changed quickly since the July 1 redeployment. Anthropic's official redeployment post said included usage on eligible plans ran through July 7; if you are reading this after that date, check your Claude account and current support notes before budgeting around any promotional access.

Link Map

What Nick Tested

The experiment was simple: ask Fable 5 to create 25 websites that demonstrate extreme web design capability, give it creative freedom, allow it to use external inspiration and media tools, require each site to be distinct, then make the agent iterate before approving the outputs.

In the transcript, Nick says the run produced sites with 3D particle systems, shaders, audio-reactive visuals, a small game, generated product visuals, experimental typography, and motion-heavy experiences. He also says the work took roughly 20 to 25 minutes, with parallel sub-agents bringing the average to about a minute per site.

The interesting part is not the raw count. Plenty of generated outputs can look impressive in a short demo and fall apart under real QA. The interesting part is the shape of the work:

• One parent goal instead of 25 tiny prompts.
• Creative freedom inside a bounded objective.
• Access to visual tools and references.
• Multiple iteration passes before approval.
• Browser verification instead of trusting the first render.
• Deployment and a guide route as part of the output.

The Prompt Is the Product

Nick's best line is the practical one: the job is not to design one site. The job is to design the prompt that designs many sites.

That is a real shift in AI web design. A normal design prompt says, "Make me a landing page for X." A stronger agentic prompt says:

• Here is the audience and reason for the run.
• Here is what variety means.
• Here are the tools you may use.
• Here are the quality bars.
• Here is how many review passes are required.
• Here is where the results must be hosted.
• Here is how to package the work for reuse.
• Do not interrupt me unless there is a real blocker.

That is not just prompting. It is workflow design. You are setting the agent's objective, constraints, tool access, verification loop, and delivery format.

The Tool Stack

The video mentions a stack that looks very different from a traditional website workflow.

The Verification Loop

The most important operational detail in the video is verification. Nick says he used Chrome DevTools MCP because he was working from a terminal view and needed a way for the agent to inspect its own web output. Anthropic's own Claude Code page makes the same broad point: Claude Code works with command-line tools and MCP servers, and Chrome's MCP documentation explains why browser inspection matters for coding agents.

For production work, I would not accept "it looks cool" as the end state. A better verification loop for AI-generated sites is:

1. Open every page in a real browser.
2. Check console errors.
3. Check mobile layout at narrow, medium, and desktop widths.
4. Run performance checks, especially for shaders, video, canvas, and large media.
5. Run accessibility checks for contrast, keyboard focus, headings, labels, and reduced motion.
6. Confirm links, forms, analytics, and deployment settings.
7. Write a short QA report before calling the concept client-ready.

This is where the human still matters. The agent can generate 25 directions. A builder has to decide which direction is strategically useful, technically safe, and commercially appropriate.

Commercial Reality

The video uses the "$10K website" frame. I understand the point: Fable can produce visual concepts that would have felt expensive or time-consuming a year ago. But a $10K client website is not just a pretty front-end experiment.

What clients actually pay for is the complete system:

• Positioning and offer clarity.
• Information architecture.
• Copy that sells without sounding generic.
• Brand fit and design consistency.
• Responsive implementation.
• Performance and accessibility.
• Conversion tracking and analytics.
• Forms, CRM, email, payments, or booking integrations.
• Launch support and maintenance.
• Judgment about what not to ship.

So I would not sell raw Fable outputs as finished premium websites. I would use them as a concept engine. Generate 10 to 25 bold directions, pick the best two, rebuild the chosen direction into a maintainable system, and sell the strategy plus implementation.

How To Adapt It

If you want to try this safely, do not start with 25 sites. Start with three.

Use one niche and one real business type. For example:

• A premium architecture studio.
• A local dental clinic.
• A SaaS analytics product.
• A boutique fitness coach.
• An AI automation consultancy.

Then ask for three meaningfully different directions:

• One conservative and conversion-focused.
• One editorial and brand-led.
• One experimental and motion-led.

That gives you range without drowning in outputs. The goal is not infinite templates. The goal is better taste, faster exploration, and a stronger client conversation.

Prompt Template

This is an original JQ AI SYSTEMS adaptation inspired by the pattern in Nick's video. It is not a verbatim transcript prompt.

You are my senior web design agent.

Goal:
Create 3 distinct website concepts for [business / niche] that show high-end web design taste, clear conversion thinking, and technical polish.

Audience:
[Describe target buyer]

Offer:
[Describe service/product]

Creative range:
1. Direction A: conservative, clear, conversion-focused
2. Direction B: editorial, brand-led, elegant
3. Direction C: experimental, motion-led, memorable

Tools you may use:
- local files and code
- browser preview
- browser inspection / DevTools if available
- approved image or video generation tools if available
- approved deployment preview if available

Rules:
- Do not copy any reference site or asset.
- Use generated or placeholder assets unless I provide approved brand assets.
- Keep each concept maintainable.
- Avoid heavy motion unless it supports the message.
- Respect accessibility, mobile layout, and performance.

For each concept:
1. Define the creative idea in one paragraph.
2. Build the prototype.
3. Run three review passes:
   - design and hierarchy
   - mobile and accessibility
   - performance and technical errors
4. Fix the issues you find.
5. Produce a short guide explaining the direction, files, dependencies, and how to customize it.

Stop only when:
- all 3 concepts render locally,
- there are no obvious console errors,
- the guide is written,
- and you have a short recommendation for which concept is most commercially useful.

For client work, add one more line: "Do not deploy publicly until I approve the final version." That small sentence can save a lot of cleanup.

Builder Checklist

• Use Fable only for the creative and agentic work that justifies the cost.
• Start with 3 concepts before scaling to 25.
• Define audience, offer, and business goal before asking for visual fireworks.
• Give the agent safe tools and explicit boundaries.
• Require review passes, not just generation.
• Use browser verification for rendered output.
• Check mobile, accessibility, performance, and console errors.
• Review generated assets for rights, brand fit, and quality.
• Turn the best direction into a maintainable system.
• Sell the full delivery layer, not just the screenshot.

Sources

• Nick Saraev: Fable 5 Is Back. Use It To Print With These $10K Websites
• Nick Saraev on X
• Anthropic: Redeploying Fable 5
• Anthropic: Claude Fable 5 and Claude Mythos 5
• Claude Platform release notes
• Claude Platform pricing
• Claude Code product page
• Claude Code /goal documentation
• Claude Code subagents documentation
• Chrome for Developers: Chrome DevTools MCP
• Chrome DevTools MCP GitHub repo
• Higgsfield Claude MCP page
• Netlify deploy overview
• Netlify create deploys