AgentSkillsCN

distill

将机器学习/强化学习的研究代码库提炼为可落地的入门指南。当您克隆某个代码仓库并希望深入理解其具体实现、发掘新颖的模式，或为未来的开发工作做好背景铺垫时，可使用此技能。该技能会生成一份 CODEBASE_PRIMER.md 文件，详细记录架构设计、算法原理、关键代码片段，以及后续适配与优化建议。特别针对机器人技术与强化学习领域的论文进行了优化。

SKILL.md
--- frontmatter
name: distill
description: Analyze and distill ML/RL research codebases into actionable primers. Use when exploring a cloned repository to understand its implementation, extract novel patterns, and prepare context for future development. Produces a CODEBASE_PRIMER.md documenting architecture, algorithms, key code snippets, and adaptation guidance. Optimized for robotics and reinforcement learning papers.
context: fork
agent: Explore
allowed-tools: Read, Grep, Glob

Distill

Analyze an ML/RL research codebase and produce a distilled primer for learning and future reference.

Workflow

  1. •Check for optional PAPER_CONTEXT.md (abstract, key contributions, relevant equations)
  2. •Explore the codebase structure thoroughly
  3. •Extract and analyze relevant components
  4. •Generate CODEBASE_PRIMER.md in the repo root

What to Extract

Focus exclusively on the research contribution. Identify and document:

  • •Model architecture definitions (networks, modules, custom layers)
  • •Loss functions and objectives
  • •Reward shaping and reward engineering
  • •Training loop logic and update rules
  • •Key hyperparameters (especially those mentioned in paper or that deviate from defaults)
  • •Novel data augmentation or preprocessing
  • •Observation and action space design
  • •Environment wrappers and modifications

What to Ignore

Skip implementation details unrelated to the core contribution:

  • •Logging and experiment tracking (wandb, tensorboard callbacks)
  • •Config/CLI parsing (hydra, argparse, omegaconf boilerplate)
  • •Checkpoint saving/loading mechanics
  • •Distributed training setup
  • •Standard dataset loading unless novel

Note these exist only if needed to run the code.

Distinguishing Novel vs Standard

Explicitly identify what comes from external libraries versus what the paper contributes:

code
They use SAC from stable-baselines3 but modify the critic network:
[code snippet of modification]

Common RL frameworks to recognize: stable-baselines3, CleanRL, rllab, Tianshou, RLlib, SKRL. Common simulators: MuJoCo, Isaac Gym, Isaac Lab, PyBullet, Gymnasium, robosuite, dm_control.

Code Snippet Format

Include actual code for novel/interesting components. Annotate with location and explanation:

[Component Name]

[One sentence: what this does and why it matters]

python
# from path/to/file.py:XX-YY
[relevant code, trimmed to essential lines]

[Explanation of the approach and design choices]

Adaptation notes: [What to consider when applying this to a different task/robot]

Keep snippets focused. Extract only the lines that matter, not entire classes.

Output Structure

Generate CODEBASE_PRIMER.md following the template in references/primer-template.md.

Paper Context (Optional)

If PAPER_CONTEXT.md exists in repo root, use it to:

  • •Map code sections to paper sections where possible
  • •Verify hyperparameters match paper claims
  • •Anchor explanations to paper terminology

If absent, proceed without. The primer should stand alone.

Final Output

Save primer to CODEBASE_PRIMER.md in the repo root. This file is designed to be used as context in a fresh conversation for applying the learned patterns.