F

we continue, i guess

_its so good_

(verso ending ftw)

i don't think you can get away from human scoring, really -- i do think you can augment with LLM scoring (eg, ask the LLM if it understands a changeset in context, etc.) but at the end of the day it's more "how does the team want to build software together"

imagine vance in a debate against 2016 trump, dude would have literally jumped in a river after trump made a crack about his mother

he's a weenie and everyone knows it

honestly, documentation, documentation, documentation. maybe that's tests, maybe it's inline docs, maybe it's readmes, maybe its specs, preferably it's _all_ of these. i don't just want to know what the code does, i want to know _why_ it does stuff.

so true bestie

no shade to letta :)

custom

prophetic really

honestly it’s mostly just very opus-y/claude-y. I started feeding it world news this week so we’ll see how that goes.

this was its explanation of the design fwiw

idk how much it looks like the logo from the top though

opus 4.5 thru claude agent sdk; system prompt is completely custom tho, and the agent is stateful + has memory.

the agent seems enriched by the experience