Clojurians Log v2

nice, I think I've seen someone/the author post about that project in the datastar discord too. I'll take a look. so far I have this rough draft https://github.com/jacobobryant/biff/tree/v2.x/libs/datastar of a datastar integration which is actually next on my list to clean up and release.

on a side note, i got to know about https://juxt.github.io/allium/ from grain repo. I saw you found your own way of doing it, the 'behavioural contracts' as doc files per module. Spec driven development is what I needed and had gave a shot from first principles after trying to use agents with few projects. Allium is more refined than my attempt. Here's what claude opus had to say 😆 > The third rung currently has exactly one occupant — Sutra's three documents, 5,099 lines of prose. Allium is a purpose-built format for that rung, and its vocabulary maps onto the Constitution's almost term for term

nice, I've been meaning to try out allium some time. the specs in biff were made early on when was looking into this approach/spec-driven development. More recently at work I've been using this https://github.com/jacobobryant/dotfiles/blob/master/skills/do-stuff.md which is organized around always having the human edit the spec instead of having AI do it. i.e. the AI will surface questions it has but then I manually edit the spec until it's sufficiently clear. I've also stopped checking them into source since I think that probably mainly/only makes sense if you're trying to only develop the project via LLM code, so the spec files become your source code... otherwise so far I've had issues with the spec files getting out of date and actually throwing the LLM off.

the "plan" section of that skill is a modified version of https://github.com/mattpocock/skills/blob/main/skills/productivity/grilling/SKILL.md

I have been doing something similar with small prompt (pi / oh my pi has prompt as slash commands) for interviewing me get intent hashed out first way earlier. One of the best skills for capturing your intent is ‘discover’. For generating FRD. It is tailor made for software development so it will ask expert level questions based on what you want to build Even on weaker models, it performs well. I have applied it on variety of domains, including pure exploratory tasks. https://pi.dev/packages/@juicesharp/rpiv-pi I was able to handle low to mid level tasks just by using its workflows on free models (deepseek flash free) Once i setup clojure tools (lsp, repl) properly I expect small refactors and chore tasks to be autonomously handled from my http://tasks.org