alchaincyf/darwin-skill
Developer toolsThis page was translated automatically. Technical details may differ from the original. View the original
Darwin.skill — A system that lets your skills evolve infinitely: evaluation→ improvement→ testing→ retention or rollback | Autoresearch-inspired autonomous skill optimization for Claude Code. Evaluate, improve, test, keep or revert.
Overview
Inspired by Andrej Karpathy's autoresearch, the autonomous experimental cycle moves from model training to the field of skill optimization. A ratchet that can only move forward. · Updated on 2026-05-28 · Absorbed two papers from Microsoft Research, SkillLens and SkillOpt, for a systematic upgrade. On June 3, 2026, Microsoft wrote in an update to the SkillOpt repository: *"gbrain, gbrain-evals, and have all integrated SkillOpt."* We absorbed its validation-gated framework, which included Darwin in its own integration list. This is a mutual tribute. 👉 Go check out the SkillOpt repository 2.0 is not patchwork but a systematic upgrade after absorbing two papers from Microsoft Research on 2026-05-22. Five changes: - The original "error handling" dimension has been upgraded to (Failure Mechanism Encoding): not just "telling agents not to make mistakes," but explicitly encoding known failure paths into skills. - The original "Clarity" dimension has been upgraded to (Actionable Specificity): Ambiguous words like "suggest/can consider/situation-based/flexible/depending on circumstances" are explicitly prohibited. - Added ninth dimension (High-Risk Action Blacklist): rm/git reset --hard/force push and other destructive operations must be explicitly banned in skills - Multiple independent judges review: 2 independent judges are activated each round - No reuse of judges: New judges are launched in the next round to avoid anchoring effects - Early stop mechanism: 30% increase in a single round is automatically alerted - Phase 1 Baseline evaluation: Automatic + manual review of reports, decide what to change - Phase 2 Single-dimensional optimization: 🔴 CHECKPOINT forced pause, wait for user confirmation - Phase 2.5 Test prompt running (optional) - Phase 3 Regression test: 🛑 STOP Forced stop when price increases fall below threshold 1. Modify and evaluate the same AI (SkillLens empirical evidence: LLM self-assessment accuracy is only 46.4%) 2. Use git reset --hard as a rollback method (apply git revert) 3. Stack redundancy just to score 4. Skip test prompts and score directly 5. Change multiple dimensions in one round 6. Dry run ratio > 30% 7. Silently skip exceptions 8. Ignore dimension-related clusters
Overview (machine translated)
English | Chinese
The animation is produced by huashu-design skill
Darwin.skill 2.0
**Optimize your Agent Skills just like training models. **
Inspired by Andrej Karpathy’s autoresearch, the autonomous experimental cycle moves from model training to the field of skill optimization. A ratchet that can only move forward.
v2.0 · Updated on 2026-05-28 · Absorbed two papers from Microsoft Research, SkillLens and SkillOpt, for a systematic upgrade.
[! NOTE] ** 🤝 Microsoft Research has included Darwin in the official SkillOpt integration list. ** On 2026-06-03, Microsoft wrote in an update to the SkillOpt repository: 「gbrain, gbrain-evals, and darwin-skill have all integrated SkillOpt.」 We absorbed its validation-gated framework, which included Darwin in its own integration list. It was a mutual tribute. 👉 Go check out the SkillOpt repository
What’s New in 2.0
2.0 is not patchwork but a systematic upgrade after absorbing two papers from Microsoft Research on 2026-05-22. Five changes:
1. Scoring criteria: 8 vitamins → 9 vitamins (absorbing 73.8% rubric prescriptions proven by SkillLens)
- The original “error handling” dimension was upgraded to Failure Mechanism Encoding: not just “telling agents not to make mistakes,” but explicitly encoding known failure paths into skills
- The original “Clarity” dimension has been upgraded to Actionable Specificity: explicitly prohibits ambiguous terms such as “suggestion/can be considered/situation-based/flexible/depends on circumstances.”
- Added ninth-dimensional High-Risk Action Blacklist: destructive operations such as rm/git reset --hard/force push must be explicitly banned in skills
2. Alignment of validation mechanisms with SkillOpt’s validation-gated design
- Multi-judge independent review: Each round starts with 2 independent judges
- No duplicate judges: Launch a new round of judges to avoid anchoring effects
- Early stop mechanism: Automatically alerts for a single 30% increase
3. Human in the Loop Three-Layer Guard (Darwin’s core distinct from SkillOpt’s fully automated design)
- Phase 1 Baseline Evaluation: Automatic + manual review of reports, decide what to amend
- Phase 2 Single-Dimensional Optimization: 🔴 CHECKPOINT is forcibly paused to wait for user confirmation
- Phase 2.
This page was translated automatically. Technical details may differ from the original.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add alchaincyf/darwin-skill