darwin-skill 3.0 drops doc voting for blind evals of actual runs, with scripts as judge
AlchainHust · x · 2026-11-15
Alchain released darwin-skill 3.0 (6.9k stars), an autoresearch-inspired system that evolves Claude Code skills via evaluate→improve→test→keep-or-revert. Key change: stop having judges vote on SKILL.md docs — the author found doc scores don't predict real runs (a 3:0 doc win ran 33 extra tool calls live; NVIDIA research shows near-zero correlation). v3.0 blind-evals actual outputs with a script as judge, keeping the shorter version on ties. Install: npx skills add alchaincyf/darwin-skill.
More from coding & agent
- arXiv: Harness^2 — recursive agent harnessing for open-world tasks — rohanpaul_ai · 2026-11-15
- Google and Northwestern's Harness² self-improves agent harnesses without answer keys — rohanpaul_ai · 2026-11-15
- Dev builds a PowerShell AI harness for Microsoft's new Intelligent Terminal — dfinke · 2026-11-15
- snitch: a harness that watches agents in parallel and reverts bad paths — generativist · 2026-11-15
- Dev: unattended agent runs burn 99% of tokens on useless reproducibility gates — generativist · 2026-11-15
- workspace-cli: open-source Rust CLI for AI agents to drive Google Workspace — majidmanzarpour · 2026-11-15