Tech • AI • Robotics • Game
GPT-6 Astra vs Fable 5.1: A $20,000 Website in One Prompt
7/10TL;DR
A single autonomous prompt produced two usable medical spa websites, with GPT-6 Astra finishing in 1 hour 18 minutes and Claude in 2 hours 45 minutes, while final client-ready quality depended on manual post-production fixes rather than the initial build alone.
KEY POINTS
One-prompt website generation
The workflow centered on writing one detailed prompt and letting two AI systems build full websites autonomously. The target was a medical spa site designed to avoid the common generic AI look of a standard hero banner, stock imagery, and repetitive feature cards. The goal was a result comparable to sites agencies often price at $20,000 or more.
Seed framework and orchestration
The prompt was assembled with a framework called Seed, used to package research, design direction, and execution rules. An orchestration layer called Paul handled the website build process, while the prompt instructed the models to research 10 competing medical spa websites, extract strong design elements, and use browser automation through Playwright CLI to gather images and context.
Autonomous research and build instructions
The systems were given broad autonomy to research, plan, apply, and unify the final site without further intervention. They were also pointed to a markdown specification file and a website file path so the process could run as a one-shot build. Both models were launched under similar settings, described as medium effort, to keep the comparison reasonably even.
Speed difference between the models
GPT-6 Astra completed its full run first in 1 hour 18 minutes, while Claude ran into issues with sub-agents and did not finish until 2 hours 45 minutes. The timing gap became one of the clearest differences between the two systems. Faster completion also mattered because heavy usage on one build consumed premium compute capacity.
Astra’s output and strengths
The site generated by Astra was judged solid for a single-pass build, especially in a market where many med spa websites are described as outdated or poorly functioning. Its design appeared sharper, with cleaner visual treatment and a strong enough foundation to be refined later. The result was considered viable despite still needing polish.
Claude’s output and trade-offs
The Claude version delivered some good animations and acceptable placeholders, but it also introduced visual problems, including an awkward animation choice and layout overlap. Its rounded styling created a noticeably different aesthetic from Astra’s sharper presentation. Even so, the build included a complete booking section and was considered workable after edits.
Functionality mattered more than style alone
The decisive factor for client delivery was not which site looked better at first glance, but whether the copy, layout, and user flows were reliable. Both sites produced working booking functionality, though the booking tools still needed backend connection to a client CRM. Overlapping elements, cutoff text, and spacing issues remained the main barriers to handoff.
Deployment was straightforward
After generation, the files could be pushed to GitHub and deployed through Railway. The process involved selecting the repository, initializing the build, generating a temporary domain, and then attaching a custom domain through DNS settings from registrars such as Namecheap, GoDaddy, or Google. This moved the site from localhost to a live production link quickly.
Post-production determined commercial value
The most important stage came after the first build. The workflow relied on reviewing the live site, checking whether the client approved the wording and color scheme, and confirming how the client’s existing software stack would affect backend edits. That review process was presented as the difference between a rough AI draft and something a business would pay for.
Editing through screenshots and prompts
Design refinement used 21st.dev as a source of replacement animations and interface ideas. Specific components such as CTA buttons, hero backgrounds, and scroll behavior were updated by taking screenshots, pasting them into the AI tool, and prompting for targeted changes while preserving the existing brand theme. The system then pushed revisions back to GitHub, allowing redeployment of improved versions.
Template reuse for future client work
Once a strong version existed, it could be reused as a template for additional clients in the same niche. The repeatable approach was to swap in a new brand kit, adjust animations, update imagery, and fine-tune copy rather than rebuilding from scratch each time. That made the process less about one perfect first prompt and more about establishing a scalable production workflow.
CONCLUSION
The experiment showed that autonomous AI can generate a credible business website in under three hours, but the first pass is only the foundation. The real value lies in speed, iterative refinement, and turning a workable draft into a polished client-ready product.
Ask a question