I Told Claude to "Make it better" 100 Times (It Went Too Far)
Published
Should I watch this?
Skip to the real content 0:22Your full summary is ready below. Read on while we finish your watch recommendation.
Your summary is ready. The watch recommendation is next.
Worth watchingYes
The Sonnet and Opus milestone walkthroughs show useful ordering improvements giving way to feature creep and broken interactions—a concrete illustration of why forced AI iteration needs human direction and a stopping point.
The main answer
Telling an AI to “make this website better” 100 times did not produce steadily better results. Both models initially improved an intentionally bad coffee-shop site, but forced iteration eventually produced clutter, unnecessary features, and—in Opus’s case—broken interactions. The practical lesson: give AI a clear design direction, evaluate each change, and stop when further additions no longer help customers.
Key takeaways
- Early changes were useful: mobile-friendly layouts, improved imagery, ordering interfaces, and pickup information. Later changes increasingly became decoration or feature creep.
- Opus produced more, not necessarily better: roughly 21,000 lines of code versus fewer than 3,000 for Sonnet, taking about 8.5 hours versus almost two.
- The final Opus site included distracting extras such as pumpkin carving and music, alongside malfunctioning interactions and an ineffective Spanish toggle.
- Neither model reportedly searched the web for inspiration. The experiment excluded skills, plugins, memory, and reference sites, so it does not show what a deliberately guided workflow could achieve.
- The instruction to ship a change every round encouraged unnecessary modifications. This tests forced iteration—not whether AI can recognize a good stopping point when allowed to stop.
- No conversion measurements, customer testing, or systematic usability scores were presented. More code, media, or ordering prompts cannot establish business improvement.
- Costs were estimates of API-equivalent usage, not actual API charges; the creator used a subscription. Opus was approximately $137, while Sonnet was inconsistently reported as $29.93 earlier and about $22 later.
Summary
The creator repeatedly applied the same improvement prompt to two Claude Code models, identified in the transcript as Sonnet 5.5 and Opus 5.5. Each round used a fresh session while continuing work on the evolving website. Both delivered substantial early improvements, but later rounds showed diminishing returns. Sonnet remained comparatively restrained; Opus expanded into a sprawling, partially broken experience. The demonstration supports human direction and selective acceptance of changes, but its forced-change setup, subjective judgments, and lack of repeated trials limit broader conclusions about model quality.
What’s being promoted?
The clearest promotion is ScrollCraft, a free website-design skill. The creator interrupts the comparison to recommend it, says they use it on every website, highlights typography, spacing, layering, and scroll-driven animation, and directs viewers to a description link. They also showcase their AI Automation Society website as evidence of their design workflow. This fits the video's message neatly: unguided AI design struggles, while the creator offers a tool intended to supply design guidance. That suggests a possible audience-building or lead-generation purpose alongside the educational experiment. However, the transcript does not establish an email requirement, paid upsell, affiliate arrangement, or sponsorship. ScrollCraft itself was not tested in the experiment, so the results do not demonstrate its effectiveness. Glido receives an on-screen demonstration to explain Canvas/WebGL, providing incidental product exposure. Claude Code is the central product being evaluated, and 21st is mentioned as a possible component resource. Neither mention establishes a commercial relationship. The closing request for likes is explicit channel promotion, but there is insufficient evidence of a concealed commercial agenda.
