whenwords: A relative time formatting library, with no code

It is interesting how the framing "it is for LLM/it will make sense with LLM" revitalises old useful ideas.

For a long time, I was kinda obsessed with "freeing" useful common algorithms in the form of descriptions+tests (accessibility of which, I believed, might be more important than a particular implementation); I approached that several times, with different completion and a different final degree of disenchantment. I was mostly focused on natural text processing algorithms/libraries which (almost) everybody uses but (almost) nobody can implement in another language, even though (it seemed to me) the most value in those libraries is just data/test cases, and a general approach.

Some of the things I did through those years ("source" software and what I tried to do with it):

• Hunspell spellchecker resulted in spylls -- while it is fully working Python reimplementation, the idea mostly was to have Python as kind of "high-level pseudocode" to explain "how it should work", ready to reimplementation (for all I know, there was at least NodeJS reimplementation, and the author of Rust spellchecker spellbook says that my explanation helped on the initial stages of work).
• Snowball stemmer: it is its own semi-declarative language and complicated algorithm for word stemming; it is both "open" (you can use it from many languages) and "closed" (the base language is quite peculiar). So, for example, modern Lunr.js search engine for its stemming uses a clone of some abandoned JS port. I tried to experiment with Ruby DSL first, to (hopefully) then come to some language-agnostic algorithm description and data representation, but never came to anything interesting.
• LanguageTool rule-based grammar checker is a HUGE software package in Java, which is actually 90% declarative rules; again, I thought about its "freeing" (into a language-agnostic form) for many years, but eventually I understood that LLM-based grammar checkers have already taken over the world, so I never published any of my experiments

...and so on, there are tons of "classic" (or not so classic) software, which is mostly "data + relatively simple algorithms", but buried in the layers of implementation.

LLM or no LLM, I believe it is fruiful to have such software "ported" into clean language-agnostic test cases + concise, straightforward descriptions.

I submitted this as a reaction to a story from five months ago, "When vibe coding, isn't the source code the prompt?" The consensus was "eh, not really." But that was five months ago and as usual, the LLM landscape shifted, and here we go, a library were the prompt is the code.

Now, perhaps I think too much about this, but this seems to fly against the current orthodoxy of "code is a liability" (no code is preferred, then code you didn't write, and only as a last resort, write code---or is the zeitgeist changing?), and even against some of the conversations I've been involved in on this site when I bring up "prompts as source code." But here we are, a repository with a single prompt that says it's "cut-n-paste."

Also, I linked directly to the repo, but there is a blog post from the author about the library.