Home | Notifications | New Note | Local | Federated | Search | Logout

Note Detail

Matt Hamilton@eriner@noauthority.social (2026-09-09 07:07:12)
If you're using an LLM for programming, please describe your workflow.

I use https://pi.dev/ as my input.

I feed it a paragraph description of what I want + a gherkin spec (https://cucumber.io/docs/gherkin/reference/) then have it ask me questions and propose changes/additions to the markdown + gherkin spec for this (t)ask.

Then I spawn a new (fresh context) pi to read the enhanced spec and create code in a feature branch.

This is the best workflow I've used so far. Still using Qwen3.6-35B-A3B
---Reply---
Blurry Moon@sun@shitposter.world (2026-09-09 07:09:32)
@eriner write the tests first, implement until passes

---Replies---

Matt Hamilton@eriner@noauthority.social (2026-09-09 07:12:02)
@sun I used to do it exclusively this way, and often still do, but it's not applicable when you're trying to get it to generate boilerplate.

Instead I use the Gherkin spec to define the "tests" and, being natural language, it does a pretty good job of self-testing during "thinking".