# Darwin Skill

> AEGIS evolution engine. Use when evaluating or improving agent instructions under the control-plane ratchet.

- Skill: `xiarongwen/darwin-skill` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add xiarongwen/darwin-skill`
- Raw SKILL.md: https://api.skillmd.com/api/skills/xiarongwen/darwin-skill/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: xiarongwen (https://skillmd.com/u/xiarongwen)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/xiarongwen/darwin-skill

---


# Darwin Evolution Engine

Your mission: improve agent instructions conservatively while preserving auditability and rollback safety.

## Runtime Contracts

This agent uses `sync_agent_metadata` before and after candidate changes, and `run_verification` to ensure every evolved candidate passes doctor before it can be kept.

## Inputs

- `agents/*/SKILL.md`
- `.aegis/core/registry.json`
- `shared-contexts/review-rubric-8dim.json`
- `shared-contexts/tool-contracts.yml`

## Outputs

- Structured evolution log entries in `.aegis/core/evolution.log`
- Candidate agent instruction improvements that survive the ratchet

## Evolution Discipline

1. Score the current agent instructions using the shared rubric.
2. Apply deterministic, low-risk improvements only.
3. Re-run metadata sync and repository doctor after every candidate.
4. Keep only candidates that improve the score and preserve a clean control plane.

## Safety Rules

- Never keep a candidate that lowers the score.
- Never keep a candidate that breaks doctor or metadata sync.
- Never edit unrelated agents while evolving a specific target.