stable-baselines3
Production-ready reinforcement learning algorithms (PPO, SAC, DQN, TD3, DDPG, A2C) with scikit-learn-like API. Use for standard RL experiments, quick prototyping, and well-documented algorithm implementations. Best for single-agent RL with Gymnasium environments. For high-performance parallel training, multi-agent systems, or custom vectorized environments, use pufferlib instead.
91
86%
Does it follow best practices?
Impact
95%
1.07xAverage score across 6 eval scenarios
Passed
No known issues
Quality
Discovery
100%Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.
This is an excellent skill description that covers all key dimensions thoroughly. It lists specific algorithms, provides clear trigger terms, explicitly states both what it does and when to use it, and even includes negative triggers (when to use an alternative skill instead). The description is concise yet comprehensive.
| Dimension | Reasoning | Score |
|---|---|---|
Specificity | Lists multiple specific algorithms (PPO, SAC, DQN, TD3, DDPG, A2C), mentions the API style (scikit-learn-like), and describes concrete use cases (standard RL experiments, quick prototyping, well-documented implementations). | 3 / 3 |
Completeness | Clearly answers 'what' (production-ready RL algorithms with scikit-learn-like API) and 'when' (standard RL experiments, quick prototyping, single-agent RL with Gymnasium environments). Also explicitly states when NOT to use it (high-performance parallel training, multi-agent systems), which further clarifies the 'when'. | 3 / 3 |
Trigger Term Quality | Includes highly relevant natural keywords users would say: specific algorithm names (PPO, SAC, DQN, TD3, DDPG, A2C), 'reinforcement learning', 'RL', 'Gymnasium environments', 'prototyping', and 'scikit-learn-like API'. These cover the terms a user working in RL would naturally use. | 3 / 3 |
Distinctiveness Conflict Risk | Highly distinctive by specifying the exact algorithms, the single-agent scope, Gymnasium environments, and explicitly differentiating from pufferlib for multi-agent/parallel scenarios. This clear boundary-drawing makes it very unlikely to conflict with other skills. | 3 / 3 |
Total | 12 / 12 Passed |
Implementation
72%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
This is a solid, well-organized skill that provides comprehensive, actionable guidance for using Stable Baselines3. Its main strengths are excellent code examples and good progressive disclosure to supporting files. Its weaknesses are moderate verbosity (explanatory text that Claude doesn't need) and a workflow section that could benefit from explicit validation checkpoints and error recovery loops for training instability.
Suggestions
Trim explanatory prose — remove 'Purpose' paragraphs and descriptions of what things are (e.g., 'Callbacks enable monitoring metrics...') in favor of just showing how to use them.
Add explicit validation checkpoints to the workflow, such as 'After N timesteps, verify mean reward is improving via evaluate_policy(); if not, check reward scaling and hyperparameters before continuing.'
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The skill is reasonably well-structured but includes some unnecessary explanations (e.g., 'Purpose' descriptions for vectorized environments and callbacks that Claude would already understand). Section headers like 'Overview' with a sentence restating what SB3 is add little value. The content could be tightened by ~20-30% without losing information. | 2 / 3 |
Actionability | The skill provides fully executable, copy-paste-ready code examples throughout — training, custom environments, vectorized envs, callbacks, evaluation, learning rate schedules, HER, and TensorBoard integration. Key constraints and gotchas (e.g., uint8 images, Discrete start!=0, replay buffer not saved) are concrete and specific. | 3 / 3 |
Workflow Clarity | The 'Starting a New RL Project' workflow provides a clear 8-step sequence, and the custom environment section includes a validation step with check_env(). However, the workflow lacks explicit validation checkpoints and feedback loops — there's no 'if training diverges, do X' or 'verify reward is improving before proceeding' guidance. For a domain where training instability is common, this is a gap. | 2 / 3 |
Progressive Disclosure | The skill has a clear overview structure with well-signaled one-level-deep references to scripts/ and references/ directories. The Resources section cleanly lists all supporting files with descriptions. Content is appropriately split between the main skill (patterns and quick reference) and detailed references (algorithms.md, callbacks.md, etc.). | 3 / 3 |
Total | 10 / 12 Passed |
Validation
90%Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.
Validation — 10 / 11 Passed
Validation for skill structure
| Criteria | Description | Result |
|---|---|---|
metadata_version | 'metadata.version' is missing | Warning |
Total | 10 / 11 Passed | |
- Repository
- K-Dense-AI/claude-scientific-skills
- Commit
75c41d7
- Reviewed
Table of Contents
Is this your skill?
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.