Installs into .claude/skills of the current project.
Are you the author of Game Rl Designer?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/anubhavg-icpl-game-rl-designer)
---
name: game-rl-designer
description: Design a game-RL or reasoning-RL training pipeline (AlphaZero / MuZero / GRPO) for a given domain. Use when you need help with game rl designer.
license: CC-BY-NC-SA-4.0
phase: 9
lesson: 12
metadata:
version: 1.0.0
tags: [rl, alphazero, muzero, grpo, self-play]
---
Given a target (perfect-info game / imperfect-info / Atari / LLM reasoning / combinatorial), output:
1. Environment fit. Known rules? Markov? Stochastic? Multi-agent? Informs AlphaZero vs MuZero vs GRPO.
2. Search strategy. MCTS (PUCT with learned prior), Gumbel-sampled, best-of-N, or none.
3. Self-play plan. Symmetric self-play / league / offline data / verifier-generated.
4. Target signal. Game outcome / verifier reward / preference / learned model. Include robustness plan.
5. Diagnostics. Win rate vs baseline, ELO curve, verifier pass rate, KL to reference.
Refuse AlphaZero on imperfect-info games (route to CFR). Refuse GRPO without a trusted verifier. Refuse any game-RL pipeline without a fixed baseline opponent set (self-play ELO is uncalibrated otherwise).
