Cucumber and AI

[honzajavorek]

Lately I thought a bit about all the AI buzz and I thought - hey, maybe, maybe we could stop writing code and just write some plain English to produce software. But the only way would be to write it in form of "tests", which the AI could use as a guidance on how to produce the code, and verify it does what it should. The "tests" are the business logic, the aim of the human. And the closest to this is Cucumber, in my opinion. This article https://danieldelaney.net/chat/ is trending on HN right now and I think it provides a similar perspective. Folks around Cucumber, what are your thoughts on this?

(I'm not looking for an answer, I'm just curious.)

[terryso]

I created a BDD demo using Cursor integrated with Cucumber. I only need to define the user stories, and then Cursor can help me generate Cucumber features based on my written user stories. Subsequently, Cursor will automatically define the step code based on the features and run BDD tests automatically, continuing until all tests pass.

https://github.com/terryso/BDD_WITH_CURSOR/blob/master/README.en.md

[davidjgoss]

Maintainer here, but only speaking for myself. Also, late - sorry.

When you're working with humans, Gherkin is a good format for describing how the software works so you have a shared understanding now, tests as you implement, and living documentation in the long term. It turns out that all translates pretty well to working with LLMs and agents - it goes best when you explain clearly what you want them to do, give them tests so they know when they're done, and more generally give them context about the codebase to avoid mistakes in future. Not only is Gherkin good for this in itself, but these models have inevitably seen a lot of it and "understand" it pretty well (euphemistically - I know they don't really understand).

I've had some success with "pairing" on the features and steps and then being hands off for the implementation. I'm definitely interested to hear what workflows have and haven't worked for others.