Everything I did in the last 7 months regarding workflows and working process with Claude.
You can reverse engineer it for yourself. Good luck.
https://t.co/XDf857C0q2
@eliautobot Three states: pass, fail, not verified, and "not verified" never opens the gate. Learned it when our suite said "0 failed" but exited non-zero. Now exit code and report must agree, and we test that every check can actually go red.
Most teams running agents through a harness assume their process is fine because the tests are green.
After 6 weeks building agent work graphs, I distilled 20 questions you need to answer before you can trust that - from "can your agent approve its own work?" to "has a check ever actually failed?"
https://t.co/wblubaOpYj
@karpathy Can my startup get access to frontier Biology? We are working on Cancer Stage 2-3 treatment plans. As of now, we have 7 out 7 patients who defeated cancer with our help (mostly Brust Cancer).
We didn’t get startups plan as well, but we have a trade mark and website.
@mattpocockuk fork my solution, you wont need to fix it :)
https://t.co/Uk9LVexebl
get the best out of it and add to ur product u sell. I use it for 5 moths in my prod development, open sourced today
16/
876/876 tests. Deterministic scripts on plain files + git. Every hook fail-open with a kill-switch.
⭐ Star it, try it:
npx -y sma-framework@latest init
then /sma-start in Claude Code
https://t.co/Uk9LVexebl
15/
Fork it, rebuild it, adapt it. Every setup is different, but these processes help anyone who builds every day.
All feedback welcome: [email protected]. Report something important and I'll ship the update promptly.
That's why it's MIT.