rl-new-project

You are scaffolding a new Python reinforcement learning project using the reinforcement-learning skill.

Arguments

$ARGUMENTS

The arguments describe the environment or task (e.g. "inventory management with discrete restock actions", "2D drone hover, continuous thrust"). If no arguments were given, ask the user what they want the agent to learn — specifically the observation the agent sees, the actions it can take (discrete or continuous), and what success looks like — before scaffolding anything.

Step 1: Load the skill

Read ${CLAUDE_PLUGIN_ROOT}/skills/reinforcement-learning/SKILL.md first. Then read:

• ${CLAUDE_PLUGIN_ROOT}/skills/reinforcement-learning/references/environments.md — custom env template, validation checklist, wrappers, and the environment testing script.
• ${CLAUDE_PLUGIN_ROOT}/skills/reinforcement-learning/references/algorithms.md — algorithm deep-dives, for justifying the algorithm choice.
• ${CLAUDE_PLUGIN_ROOT}/skills/reinforcement-learning/references/training.md — hyperparameter starting points, reward engineering, normalization.
• ${CLAUDE_PLUGIN_ROOT}/skills/reinforcement-learning/references/debugging.md — the pre-training sanity checklist to turn into a validation script.

Step 2: Choose the algorithm

Apply the SKILL.md Algorithm Decision Tree to the task description: action space type first (discrete vs. continuous), then sample-efficiency and stability requirements. State the chosen algorithm and cite the decision-tree branch and the relevant section of references/algorithms.md that justifies it. Default to PPO when the task description leaves the tradeoffs ambiguous — it is the stable choice per the skill's quick-selection table.

Step 3: Generate the project

Create this layout in the current directory (ask before writing if the directory is not empty):

<project>/
├── pyproject.toml            # or requirements.txt — gymnasium>=1.0, stable-baselines3>=2.0, tensorboard, optuna
├── Makefile
├── src/
│   ├── __init__.py
│   ├── env.py                # custom Gymnasium environment
│   ├── train.py              # SB3 training entry point
│   └── tune.py               # Optuna tuning stub
├── scripts/
│   ├── check_env.py          # Gymnasium API validation
│   └── sanity_checks.py      # pre-training sanity checklist
└── models/                   # checkpoints + VecNormalize stats (gitignored)