菜单
Apply HyperAgents' self-referential improvement pattern to any code artifact. Triggers when Claude is asked to 'improve', 'optimize', 'evolve', or 'self-improve' code, agents, skills, or prompts. Also triggers on repeated failures as an automatic recovery strategy.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Create and configure domain-specific evaluation harnesses for the HyperAgents evolution loop. Defines how tasks are loaded, agents are invoked, predictions are collected, and scores are computed. Triggers when setting up evaluation domains or creating custom fitness functions.
Manage the HyperAgents evolutionary archive — an append-only log of all code generations with fitness scores, lineage tracking, and diff storage. Triggers when working with .hyperagents/ directory, archive.jsonl files, or generation metadata.
Domain-agnostic fitness evaluation for evolved code generations. Defines evaluation harness interfaces, scoring contracts, and multi-domain aggregation. Triggers when evaluating code quality, running benchmarks, or scoring agent outputs.
Evolutionary parent selection algorithms for choosing which generation to mutate next. Implements random, best, score-proportional, and novelty-aware selection. Triggers when selecting parents, managing exploration/exploitation tradeoffs, or configuring evolution strategy.
Two-phase evaluation strategy from HyperAgents — run a quick staged check on small samples first, only proceed to full evaluation if the staged eval passes. Saves 90%+ compute on broken mutations. Triggers when evaluating generations, running benchmarks, or optimizing evaluation cost.
| name | Self-Referential Self-Improvement |
| description | Apply HyperAgents' self-referential improvement pattern to any code artifact. Triggers when Claude is asked to 'improve', 'optimize', 'evolve', or 'self-improve' code, agents, skills, or prompts. Also triggers on repeated failures as an automatic recovery strategy. |
| version | 1.0.0 |
| metadata | {"filePattern":["**/*.md","**/SKILL.md","**/agents/*.md","**/.hyperagents/**"],"bashPattern":["hyperagents","evolve","self-improve"],"priority":90} |
This skill implements the core insight from Meta's HyperAgents paper: agents can improve themselves by treating their own code/prompts as mutable targets, evaluating changes with fitness functions, and selecting the best mutations across generations.
/hyperagents:evolve┌─────────────────────────────────────────────┐
│ GENERATE LOOP │
│ │
│ 1. SELECT PARENT (from archive) │
│ │ │
│ 2. MUTATE (meta-agent modifies code) │
│ │ │
│ 3. EVALUATE (fitness scoring) │
│ │ │
│ 4. ARCHIVE (store generation + score) │
│ │ │
│ 5. SELECT NEXT PARENT (fitness-weighted) │
│ │ │
│ └──────── repeat ────────────┘ │
└─────────────────────────────────────────────┘
Unlike traditional agent frameworks where the orchestrator is fixed, HyperAgents allows the meta-agent to modify:
Changes are not accepted blindly. Every mutation is evaluated:
All generations are kept in an append-only archive:
All changes are made in isolation:
When asked to self-improve a target:
1. Create .hyperagents/ directory if not exists
2. Define fitness function for the target
3. Store initial version as gen_initial
4. For N generations:
a. Select parent from archive
b. Create git worktree
c. Dispatch meta-agent to modify target in worktree
d. Capture diff
e. Run fitness evaluation
f. Store results in archive
g. Clean up worktree
5. Apply best generation's changes