Published: 2026-10-30
Summary

Telling Claude to "make it better" 100 times: the exact prompt, a clean-room setup, and where Sonnet and Opus 5.5 started overbuilding

Chapters / key moments (click to jump — plays here on the page)

A deliberately plain site was handed to each model 100 times with an instruction to make one bold change per round. The video is a useful look at what "keep improving" does without a human deciding when to stop.

Source video

"I Told Claude to "Make it better" 100 Times (It Went Too Far)" by Nate Herk — Watch on YouTube → The speaker used his subscription and quotes API-equivalent costs. One site, one run per model.

The setup

Start from a deliberately ugly, plain site (the speaker made it awful so there was room to improve).
Run each model in a clean room every round: a fresh session with no skills, no CLAUDE.md, no plugins and no memory, so each has only its own knowledge and built-in tools.
Send the same prompt 100 times in a row.
Record per round the tools called, lines of code and cost, and mark milestones (first image, first video, first form).
Run Sonnet on medium effort and Opus on medium effort and compare.

The prompt

As read in the video:

Look at the site first, then innovate. Think about the user experience and think outside the box. Make one bold change that improves conversions or the look and feel, not a small tweak. Use any tool you have. You must ship a change every time.

The last line was needed because in earlier tries the model stopped around round 20 to 25 and said nothing more could be done.

What happened

Sonnet: ran nearly two hours. Cost reads as about $29.93 at one point and about $22 in the final summary, so treat it as roughly $22 to $30. It made 11 images and two videos, first video in round 24, first form around round 32, and ended under 3,000 lines of code. Between rounds 32 and 100 nothing was flagged as a milestone.
Opus 5.5: ran about 8.5 hours, about $137, with 54 AI images and three videos, first video in round 7, first form in round 15 (including a QR code that texts an order), first canvas/WebGL element in round 32, and about 21,000 lines of code.
Lines of code climbed steadily while cost per round moved up and down, so code was added every round without evidence the site was getting better.
Round 100 was worse than a middle round. Sonnet had removed a feature it earlier added and changed the font late on. Opus's final site was cluttered (falling leaves, a pumpkin-carving widget, caffeine tracking, a music player, a language toggle that did not switch the language) and some sections no longer worked.
Neither model searched the web for inspiration, even when told to innovate and think outside the box.

Caveats: the lessons

Forcing a change each round turns "innovate" into "add". The speaker stopped liking the output well before round 100 (for Sonnet about round 65; for Opus around round 32).
Both optimised hard for conversions, ending in upselling and forced ordering, a hard sell rather than a soft close.
His summary: the model sets the ceiling and the prompt sets the direction. A human has to decide when to stop and what to keep.
One plain site and one run each, judged by one person; the model versions beyond "Sonnet" and Opus 5.5 were not stated. A web-design skill or reference sites would likely change the result, as the speaker notes.