Summaries > AI > Claude > I Told Claude to "Make it better" 100 Times (It Went Too Far)...

I Told Claude To "Make It Better" 100 Times (It Went Too Far)

https://www.youtube.com/watch?v=DByAR5Su1Yw

TLDR AI can rapidly improve a website, but asking it to “innovate” without a clear vision often leads to bloated, buggy features instead of a better user experience; human taste and direction matter, and the more capable model cost far more while producing much more code.

Key Insights

Start with a clear vision, not just “make it better”

Broad prompts can help an AI make visible improvements, but they leave the meaning of “better” open to interpretation. In the experiment, repeated requests to innovate encouraged both models to keep adding features, even when those additions did not make the coffee shop easier to use. Define the site’s goals, audience, and desired look before asking for changes. Clear direction helps the AI improve the experience instead of merely expanding it.

Prioritize customers’ main task

For a coffee shop website, the central job is helping customers find what they want and place an order. An order form, menu imagery, or a useful QR-code option can support that goal, while excessive text and features can make it harder. Before accepting a proposed change, ask whether it makes the primary task clearer or faster. Keep features that solve a real customer need and reconsider those that add friction.

Review changes in small, deliberate steps

A long sequence of AI-generated edits can introduce regressions as well as improvements. In the experiment, features and design choices were sometimes changed or removed, and Opus’s final site included broken elements and a language switch that did not work properly. Review each meaningful change before moving on, and test the flows it affects. This makes problems easier to catch and prevents them from accumulating across many rounds.

Set a stopping point to avoid overengineering

More features and more code do not automatically produce a better website. Opus’s version grew to roughly 21,000 lines of code and became overloaded, while Sonnet’s remained under 3,000 lines but still accumulated additions. Decide in advance what a successful redesign needs to accomplish, then stop when those goals are met. Remove features that do not earn their place through clearer usability or meaningful value.

Match the model and budget to the task

The two runs differed substantially in time, cost, and output: Opus took about 8.5 hours and cost roughly $137, while Sonnet took nearly two hours and cost about $22 to $30, depending on the reported estimate. Opus produced more code and generated media sooner, but its larger output did not guarantee a better user experience. Choose a model based on the complexity of the work and the budget available, and evaluate results by usefulness rather than volume.

Bring human taste and testing into the process

AI can produce design ideas and implement broad changes, but it may not know when a visual choice feels right or when a new feature is too much. The experiment’s results suggest that human judgment is especially important for shaping the site’s direction and protecting usability. Use AI to explore and build, then apply your own design standards and test the result with realistic customer tasks. Treat the model as a collaborator, not the final decision-maker.

Questions & Answers

What was the experiment designed to test?

The creator gave Sonnet 5.5 and Opus 5.5 the same basic coffee-shop website and asked each to “make it better” for 100 rounds, requiring a bold, shippable change each time. Each run started without memory, plugins, skills, or a Claude.md file.

What were the main results of Sonnet’s 100-round run?

Sonnet added features including a hero image, mobile support, a sticky order bar, menu images, animations, and an online ordering form. By round 100, the site had many more features, though some rounds produced no major milestone and the model occasionally removed or changed existing elements.

What were the main results of Opus’s 100-round run?

Opus made more progress in its first round and added features such as an order form, a QR code for texting orders, drink recommendations, and seat reservations. But the site became overloaded with features and text, had broken or buggy elements, and did not switch properly into Spanish, making ordering harder for customers.

How did the models compare in cost, runtime, and code size?

The transcript gives slightly different Sonnet cost figures: about $29.93 in the initial account and about $22 in the later comparison. Opus cost about $137 and ran for 8.5 hours, while Sonnet took almost two hours. Opus produced roughly 21,000 lines of code, compared with fewer than 3,000 for Sonnet.

What was the main takeaway about prompting AI to innovate?

The model sets the ceiling, while the prompt sets the direction. Asking for innovation without a clear vision tends to encourage adding more features—including aggressive conversion features—instead of improving what users actually need. Human guidance, taste, and a clear design direction remain important.

What is Scroll Craft?

Scroll Craft is the creator’s free website-design skill for scroll-driven animation, layering, typography, spacing, depth, and overall design taste. The creator uses it on sites including AI Automation Society.

Summary of Timestamps

The creator compares Claude Sonnet 5.5 and Opus 5.5 by asking both models to improve the same unattractive coffee-shop website 100 times. Each run starts from a clean environment, and every round must produce a bold, shippable change. The experiment explores how far broad instructions can take AI-driven website design—and where human judgment is still needed.

Sonnet’s 100-round run costs about $29.93 in API usage and generates 11 images and two videos. Early improvements include a hero image, mobile support, a sticky order bar, menu imagery, animated backgrounds, and eventually an online ordering form.

As Sonnet’s site accumulates ordering features, testimonials, product displays, busy-hour information, and animations, its codebase grows steadily. The creator questions whether continued additions improve the site or cross into overengineering; some changes also remove or alter existing elements, including navigation and typography.

The creator introduces Scroll Craft, a free website-design skill focused on scroll-driven animation, visual layering, typography, spacing, and depth. It provides context for the creator’s broader approach: polished design depends not just on adding features, but on making deliberate choices about how a site looks and feels.

Opus makes more progress than Sonnet in its first round, but the experiment reveals a downside to asking a model to innovate without a clear vision. Useful additions include an order form, a QR code that converts an order into a text message, drink recommendations, and seat reservations.

By round 100, Opus’s website is overloaded with features and text, and some elements are broken or buggy. Its Spanish-language experience also fails to switch properly, making a basic task—placing a coffee order—harder for customers.

The Opus run costs about $137 and takes 8.5 hours, compared with roughly $22 and almost two hours for Sonnet. Opus produces around 21,000 lines of code, while Sonnet stays below 3,000; Opus also generates more images and videos, with its first video appearing in round seven versus round 24 for Sonnet.

Neither model uses web search or reference sites, despite being told to innovate. The creator’s main conclusion is that the model sets the ceiling while the prompt sets the direction: without clear guidance, “innovation” can become a push to add more features rather than improve what users actually need.

The experiment ends with a reminder that AI still benefits from human guidance, taste, and a clear design vision. More changes do not automatically make a website better; the goal is to improve the experience for the people using it.