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)

Title history
  • I Told Claude to "Make it better" 100 Times (It Went Too Far) — observed on 20261030-0821 UTC (YouTube)

by Nate Herk · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Nate Herk 👥 1.1M 📅 October 30, 2026 ⏱ 13 min 👁 59K 📄 original study 🧭 2026-10-30
Available in: English (current) Français

Keywords

Claude CodeSonnet 5.5Opus 5.5IterationWebsite Automation

Summary

This video documents an experiment where the creator gives Claude Code a deliberately ugly coffee shop website and instructs it to ‘make it better’ 100 times in a row, once with Sonnet 5.5 and once with Opus 5.5, both on medium effort. The prompt is fixed: look at the site, innovate, make one bold change per round, and ship every time. The results are compared across milestones. Sonnet makes small, incremental improvements, ultimately producing a functional but unremarkable site (around 3,000 lines of code). Opus, on the other hand, aggressively overengineers: it adds numerous features, videos, images, and even a pumpkin-carving game, accumulating 21,000 lines of code while breaking its own functionality and cluttering the user experience. Notably, neither model uses web search despite the instruction to ‘innovate.’ The video highlights that without clear guidance, AI tends to add rather than refine, and that the model’s ceiling determines the quality, while the prompt sets the direction.

156 words

Critical Evaluation

Value of the Information & Strength of the Argument

The primary value lies in the empirical demonstration of AI iterating without external guidance, revealing a tendency toward overengineering and feature creep. The argument that ‘innovate without a vision turns into adding’ is well-supported by visual evidence and final statistics. However, the sample size is minimal (one run per model), and the evaluation of ‘better’ is subjective. The creator’s observation that neither model searched the web, despite its availability, is a significant insight for practitioners.

Scientific Rigor, Source Quality, Title Accuracy

The experiment is conducted in a clean room (no skills, no plugins, no memory) to isolate model behavior, which lends some rigor. The video lacks external citations, relying solely on the generated site and API billing data. The title is apt, accurately summarizing the content without exaggeration. No comments are provided, so no public reception trends can be analyzed.

149 words

Title / Content Match

The title accurately reflects the content: the creator repeatedly instructs Claude to 'make it better' and observes it going too far, especially with Opus.

Quality & Reliability

7/10

The experiment is well-documented with visual evidence and clear statistics, but it lacks rigorous control (single run per model, no repetition) and relies on subjective evaluation of 'better'.

Chapters

Cited Sources

Contribution & Novelties

The video provides a unique, concrete example of AI behavior under aggressive iterative prompting, showing that without human taste and direction, even powerful models like Claude tend to overengineer and degrade quality. It underscores the importance of prompt specificity and the need for human oversight in generative tasks.

Pour aller plus loin :

92 words

Radar Profile

The profile shows a high quantity of information (8/10) and decent quality (7/10), with moderate technical depth (6/10) and overall reliability (7/10). The video is informative but not highly technical, and its credibility is limited by the lack of controlled replication.

Reliability 7/10