Brian Hung
Projects
auto-distill
Turns Claude Code and Codex sessions into reusable skills, scored against the work people already did.
fitness, 0 to 1
top 3
revisions
candidate skills
6 drafts from one task
mixedScore
tests, and a rubric judge
rank by fitness
rankAll, keep the best
revise the best
alone, and pooled
held-out tasks
what the expert did
best skill
saveSkill to sqlite
The model writes every candidate; only the grounded score decides which survive.
The loop stops at 100% or after three rounds; publishing the best skill is a separate, human step.