Earlier this year I moved to a new apartment, and scattered a box of poetry magnets all over my fridge. Whenever people came to visit, I noticed how they always loved making poems of their own, not to mention enjoying the poems others had already created. It became such a fun and entertaining tradition, (especially due to the range of creative messages that my friends chose) that I felt compelled to make a full stack project out of it.
While LLM agents have made the technical aspects of building a project like this easier, it took me a really long time to nail the algorithm for choosing daily word sets. I started with a huge collection (on the order of 100,000) of words, grouped by part of speech and sorted by the frequency of use in English. I tried randomly choosing a few words in the top 10, a few in the top 100, and so on, for each part of speech.
That straight up did not work. Trying to make poetry out of those wordsets felt terrible. It was very discouraging. I had to go back to the drawing board many times. My current solution came about when I found some databases of words, one sorted by their sensory value (from most to least sensory-evoking) and one where words are sorted on a spectrum of concreteness to abstractness.
Another key to picking fun and interesting word sets was ensuring there are a solid set of words that are included in every set no matter what. This "glue" set was handpicked by words that I found myself searching for over and over again when trying to write poems, but it is deliberately small to help people be creative in how they express ideas.
I used Deepseek v4.1 Flash and the Pi agent. My AGENTS.md emphasizes my status as a junior developer, specifically prompting the agent to never make architectural decisions on my behalf, to quiz me on the architecture/code involved before implementing changes I request, and to never write copy so I can make sure the site feels warm and genuine in my own words. I wanted to avoid letting the code get out of control for when I want to add features in the future.
One of the things I love about building these kinds of projects that LLMs enable is being able to add tiny features that most people will never see. For example, I added special effects that occur when a haiku is written.
I chose TanStack Start and React this time instead of SvelteKit. The experience was about the same. And kudos on poems are deliberately spammable. I like seeing number go up.
Friends have expressed enjoyment from using it so far, but I'm excited to hear feedback on y'all's experience and see if there's anything huge I'm missing. Thank you in advance!