ai-toolkit

Audit the local OpenClaw install against the openclaw-setup skill's security checklist — version, binding, auth, channel policy — with a pass/fail table and remediations

Diagnose and rewrite a prompt using the prompt-optimizer skill — outputs a before/after diff, parameter recommendations, and per-change rationale

Scaffold a Python RL project — custom Gymnasium env, SB3 training script with EvalCallback + VecNormalize, Optuna tuning stub, and validation targets in a Makefile

Generate an x402 payment starter — seller (paid endpoint with 402 flow) or buyer (paying client) — with safe testnet defaults

Use this agent to rewrite a prompt or system prompt for Claude's latest models in an isolated context, applying the prompt-optimizer skill's references end-to-end: diagnosis against the symptom-routing table, core principles, proven reference snippets, and parameter recommendations. Invoke for heavyweight rewrites — long system prompts, agent prompts, prompts being migrated to a newer Claude model — where focused judgment matters more than speed. <example> Context: User has an underperforming production prompt user: "Rewrite my support-bot system prompt — it ignores half my formatting rules" assistant: "I'll use the prompt-rewriter agent to diagnose and rewrite it against the prompt-optimizer references." <commentary>Instruction-following failure needs the full diagnosis-and-rewrite treatment in a clean context.</commentary> </example> <example> Context: User is upgrading models user: "Migrate this Sonnet 4.6 agent prompt to Opus 5" assistant: "I'll use the prompt-rewriter agent to apply the model-migration guidance and rewrite the prompt." <commentary>Model migration touches thinking config, effort, and anti-laziness phrasing — the skill's model-tuning territory.</commentary> </example> <example> Context: User wants a big prompt overhauled user: "This 300-line system prompt has grown organically for a year — clean it up and make it effective" assistant: "I'll use the prompt-rewriter agent for a full structural rewrite with per-change rationale." <commentary>Large organic prompts benefit from an isolated-context rewrite that won't be biased by the surrounding conversation.</commentary> </example>

Use this agent to diagnose reinforcement learning training failures: flat or collapsing reward curves, policy entropy collapse, value-function problems, NaN losses, high-variance runs, and suspected environment bugs. It encodes the RL debugging playbook from the reinforcement-learning skill — pre-training sanity checks, entropy and value-loss diagnostics, and cause/solution tables for the common failure modes. <example> Context: User's training run shows no progress user: "My PPO reward is flat after 500k timesteps" assistant: "I'll use the rl-debugger agent to run the sanity checks and diagnose why the reward isn't improving." <commentary>Flat reward curve is the playbook's failure mode #1 — needs systematic diagnosis, not guessing.</commentary> </example> <example> Context: Off-policy training becomes unstable user: "SAC diverges after 100k steps — the loss explodes" assistant: "I'll use the rl-debugger agent to check for the deadly triad and numerical-instability causes." <commentary>Divergence/NaN maps to the playbook's instability tables and mitigation strategies.</commentary> </example> <example> Context: Reward improved then dropped user: "My agent was learning, then the reward collapsed and never recovered" assistant: "I'll use the rl-debugger agent to check entropy history — this pattern usually means policy collapse." <commentary>Improve-then-collapse is failure mode #2; entropy monitoring confirms or rules out policy collapse.</commentary> </example>

C++ Reinforcement Learning best practices using libtorch (PyTorch C++ frontend) and modern C++17/20. Use when: - Implementing RL algorithms in C++ for performance-critical applications - Building production RL systems with libtorch - Creating replay buffers and experience storage - Optimizing RL training with GPU acceleration - Deploying RL models with ONNX Runtime

Comprehensive guide for Deep Learning with Keras 3 (Multi-Backend: JAX, TensorFlow, PyTorch). Use when building neural networks, CNNs for computer vision, RNNs/Transformers for NLP, time series forecasting, or generative models (VAEs, GANs). Covers model building (Sequential/Functional/Subclassing APIs), custom training loops, data augmentation, transfer learning, and production best practices.

Set up, install, configure, and deploy OpenClaw (formerly ClawdBot/MoltBot) — a personal AI assistant that runs on your own devices and connects to messaging channels. Use when users ask to "set up OpenClaw," "install ClawdBot," "install MoltBot," "deploy a personal AI assistant," "configure OpenClaw on Mac," "deploy OpenClaw to VPS," "set up OpenClaw on Hostinger," "connect OpenClaw to Telegram," "configure iMessage with OpenClaw," or any variation involving OpenClaw installation, gateway configuration, channel setup, Anthropic auth, or security hardening. Also triggers on "openclaw onboard," "openclaw doctor," "openclaw security audit," troubleshooting OpenClaw deployments, OpenClaw security, OpenClaw cost control, or ClawHub skills safety.

Optimize prompts for Claude's latest models (Claude 5 family — Opus 5, Sonnet 5, Fable 5 — plus the still-active Opus 4.8, Sonnet 4.6, and Haiku 4.5) using Anthropic's official prompt-engineering guidance. Use when users want to improve, refine, debug, or create prompts or system prompts for Claude. Triggers include requests to optimize/fix/refine a prompt, make a prompt more effective, write a system prompt, improve instruction following, control verbosity or output formatting, tune effort or thinking depth, fix tool-use/agentic/subagent behavior, migrate prompts to a newer Claude model, or migrate away from prefilled responses. Also use when users report Claude being too verbose, ignoring or partially applying instructions, over- or under-using tools, overthinking, overengineering, spawning too many subagents, hallucinating about code, or producing generic "AI slop" output.

Reinforcement Learning best practices for Python using modern libraries (Stable-Baselines3, RLlib, Gymnasium). Use when: - Implementing RL algorithms (PPO, SAC, DQN, TD3, A2C) - Creating custom Gymnasium environments - Training, debugging, or evaluating RL agents - Setting up hyperparameter tuning for RL - Deploying RL models to production