The Poetry Industrial Complex

Moon viewing, no words. Now AI writes poetry— Measuring which path.

October 1, 2025

At a moon viewing festival at the Seattle Japanese Garden, a poet invited me to write a haiku.

All I could think of was: "There once was a man from Nantucket."

That moment of poetic inadequacy led to something unexpected: I wrote a prompt that writes programs that convert any input into haiku, iambic pentameter, and limerick. Next time someone asks me for an impromptu poem, I'll be ready!

But more importantly, I built each program four different ways directly from the same initial prompt—using four distinct AI development methodologies. AI judges the code quality, which means that I know which of the four outputs works the best for what I need it to do. (And for the most part, all four work!)

And then I wrote [EDIT: a prompt that writes] a script to judge the output -- the poetry, not the code that generated the poetry -- using three separate local models: Gemma, Phi, and Ollama.

While conducting all these experiments, which I'm sharing freely in the spirit of open science, I've been generating some incredibly useful findings, which you'll also find in the repo.

And right here, right now, mostly because I can, I'm sharing some of my research findings in poetic form. Forsooth, enjoy!

External Library Efficiency Just because you can automatically generate code doesn't mean that you should.

Research Summary:

External libraries outperform AI-generated implementations by ten to one thousand times. For well-solved problems like date validation, using established libraries provides better performance, fewer bugs, and less maintenance burden than custom AI-generated code. → Read full finding

External libraries far outshine, Artificial code with lesser design. Ten times their strength in every test, Validation of dates, they are best. With fewer bugs and less work each day. Complexity Matching Principle A method for every madness

Research Summary:

Effective AI-assisted development requires matching methodology complexity to task complexity. Over-engineered approaches waste time on simple tasks, while under-engineered approaches fail on complex systems. The right methodology depends entirely on problem characteristics. → Read full finding

When software doth entwine with Claude's insight, Matching complexity of methods to the test, Lest over-reach doth waste the tedious night, And under-serve, with failure's bitter guest, The right approach, born of problem's noble breast. Monte Carlo Methodology Sampling Do you think these poems generate themselves? Oh wait...

Research Summary:

Generating multiple samples and selecting the best produces twenty percent quality improvement. This Monte Carlo approach works across all methodologies, making it a practical production technique for critical code where quality matters more than development speed. → Read full finding

In Monte Carlo's realm of probabilistic might, Twenty percent, a gain in quality doth shine With samples bred and selected, chosen bright The best doth emerge, and flaws are left behind This method, across all ways, doth prove its right Clone the full repo: https://github.com/ivantohelpyou/spawn-experiments