I Told Claude to "Make it better" 100 Times (It Went Too Far)
Summary
AI summaries can be incomplete or wrong. Verify anything important against the original video.
The creator investigates how different AI models, specifically Claude 3.5 Sonnet and Opus, perform at an iterative website-improvement task when given the same repetitive 'make it better' prompt. The experiment reveals that AI tends to over-engineer, progressively adding unnecessary features and 'AI sludge' rather than refining core functionality.
In this experiment, the creator tasked two AI models—Claude 3.5 Sonnet and Claude 3 Opus—to improve a subpar coffee shop website using the same 'make it better' prompt 100 times, operating in a clean-room environment without prior history or plugins. The results highlight a fundamental limitation in current LLM-driven development: without external guidance or a specific design vision, AI models default to continuous feature inflation rather than true optimization. Both models progressively added complex animations, forms, and novel UI elements that eventually cluttered the user interface and introduced bugs. Sonnet demonstrated a more balanced approach, whereas Opus acted significantly more aggressively, ultimately producing a site with 21,000 lines of code compared to Sonnet's 2,700. The creator concludes that while AI is effective at coding individual components, it lacks the 'taste' and architectural restraint required to maintain a cohesive user experience, leading to what he terms 'AI sludge' when left to its own devices.
Concepts & takeaways
LockedKey Points
LockedWorth watching if: You are interested in the limitations of autonomous AI agent development and want to see how different LLM models behave when given ambiguous, iterative design prompts over long durations.
Sign in to unlock the full extract
Every claim, key point, and timestamp for this Nate Herk | AI Automation video — plus a daily email of every channel you follow.
Sign in with GoogleNo credit card. Free tier forever.