AgentSkillsCN

recursive-arena

每轮迭代都以多模型对抗作为生成器，持续推动递归式优化进程。当用户既希望进行递归优化，又期待模型间的激烈较量，或是渴望体验“递归竞技场”时，此技能便是理想之选。

SKILL.md
--- frontmatter
name: recursive-arena
description: Orchestrates recursive iterations using the multi-model battle as the generator each round. Use when user wants recursive refinement plus model battles, multi-LLM consensus, or says "recursive arena".
allowed-tools: Read, Bash(python:*)

Recursive-Arena Orchestrator

This skill composes two skills:

  • •multi-model: generates candidate answers by rotating models and collecting judge feedback.
  • •recursive: provides the macro-loop structure (decompose → critique → reflect → refine → converge).

Core behavior

For each recursive iteration:

  1. •Call the multi-model script to produce the best candidate for the current prompt.
  2. •Treat the multi-model best answer as the iteration’s Current Solution.
  3. •Use multi-model judge summaries as the primary input to Self-Critique and to update Reflection Memory.
  4. •If not converged, refine the prompt (or add constraints) and run the next iteration.

How to run

This orchestrator uses the multi-model runner bundled in this plugin:

bash
python3 "${CLAUDE_PLUGIN_ROOT}/skills/recursive-arena/scripts/recursive_arena.py" --prompt "<your task>" --iters 4 --arena-iters 3

Configuration

Multi-model configuration is read from .env (same rules as the multi-model skill):

  • •ARENA_MODELS, ARENA_OPENAI_BASE_URL / provider variants.

Optional orchestration env:

  • •RLM_ARENA_ARENA_ITERS default for multi-model per outer iteration
  • •RLM_ARENA_MAX_JUDGES default judge cap

Output

  • •The final answer is the best outer-iteration result.
  • •Show an evolution summary table with:
    • •Iteration number
    • •Writer model ID used by arena winner (numeric ID only)
    • •Average judge score
    • •Key refinement applied

Never disclose provider/model names; only numeric IDs. Never print secrets from .env.