Continuously Improve Generated Skills

Maintains and evolves perspective skills after initial distillation. Darwin.skill uses automated experimentation loops, 8-dimensional quality scoring (clarity, applicability, edge-case resilience, etc.), a ratcheting mechanism that only preserves improvements, and independent sub-agent evaluation to prevent drift. Applied to Nuwa at Phase 5 during initial generation, and can run continuously on deployed skills to incorporate new research or correct underperforming models.

1. Install Darwin.skill alongside Nuwa: npx skills add alchaincyf/darwin-skill. 2. Run autonomous optimization on an existing perspective skill (Phase 5 happens automatically during Nuwa generation, or invoke manually: darwin improve [skill-name]). 3. Darwin spawns sub-agents to evaluate the skill across 8 dimensions, test against diverse prompts, score improvements. 4. Any iteration that improves net score is kept; degradations are automatically rolled back. Output: iteratively refined skills that maintain backward compatibility while improving specificity and resilience.

Darwin.skill runs as a companion process and requires the target skill to have a properly structured SKILL.md with frontmatter (name, description, trigger_words). The improvement loop is async and can run on a schedule (hourly, daily, weekly) without blocking skill availability. Integrates with the research files in references/ directory to validate new extractions against source material. Requires Python 3.8+ and agent-based testing capability (same as Phase 4 validation).