Lev Lazinskiy

Generated Dependencies

The other day I wrote this chunker for my Bluesky dagger module using Claude that automatically splits up long posts into chunks and posts the correct amount of things onto Bluesky. It took a few seconds to generate and worked on the first try. Then I had this brilliant idea to add a few tests to make sure it keeps working for the long haul and all hell broke loose.

The dagger module and chunker were written in typescript. Claude tried to write some tests and run them using Jest. It failed because of ESM, or commonJS, or BDSM, or whatever the hell, and it took hours and hours to debug. At the end I felt completely defeated. Then I had another brilliant idea.

“Claude, make me a test runner with no dependencies to test out this chunker you made”

It did it in a few seconds, single shot and worked perfectly.

This got me thinking that we are now basically at the spot where there really is no difference between grabbing some library off the shelf and shoving it into my project, or having the llm generate one on the fly. I can see myself doing the latter more and more because it results in simpler code that both me and the llm can understand. I don’t need every feature of Jest, I needed the test cases to run, the assertions to be evaluated, and a nice green checkmark to appear.

Armin Ronacher recently wrote about this as well in the context of dependency churn. The latest llms make it easier than ever to just grab or create the functionality you need rather than signing up for the endless dependency churn and maintenance caused by many libraries.