Create, update, or organize Claude Code skills in this repo. Use when adding a new skill, reviewing existing skills for consistency, or maintaining the skill taxonomy.
中文简介正在准备中。当前显示原文说明。
实测评分 未实测
查看安装方法安装前可以查看交给 AI 代理的指令内容。
This meta-skill governs how skills are created and maintained in the tinker-cookbook repo.
All skills in .claude/skills/ are organized into 5 layers:
setup, models, hyperparams, logging)Scope: Getting started, model selection, hyperparameter guidance, training output analysis. Cross-cutting concerns needed before touching any code.
Auto-invocation: Yes — triggers when users ask about setup, models, hyperparameters, or debugging.
Key principle: These inform all other layers. Reference docs/, README.md, tinker_cookbook/hyperparam_utils.py.
tinker-sdk, tinker-types, tinker-cli)Scope: Raw Tinker Python SDK APIs — ServiceClient, TrainingClient, SamplingClient, RestClient, types, errors, and CLI commands.
Auto-invocation: Yes — triggers when users ask about Tinker API basics or CLI usage.
Key principle: Reference docs/api-reference/ for authoritative API docs.
renderers, environments, weights, completers, checkpoints, evals, datasets)Scope: Building blocks in tinker_cookbook/ — renderers, RL environments, weight lifecycle, completers, checkpointing, evaluators, dataset construction.
Auto-invocation: Yes — triggers when users ask about specific primitives.
Key principle: Reference source code in tinker_cookbook/ and docs in docs/.
sft, grpo, distillation, dpo, rlhf, multiturn-rl)Scope: End-to-end training workflows built on Layer 1 + Layer 2.
Auto-invocation: Yes — triggers when users want to set up a specific training method.
Key principle: Reference recipes in tinker_cookbook/recipes/ and defer primitive details to Layer 2 skills.
new-recipe, ci, contributing, manage-skills)Scope: Development workflow — scaffolding, testing, CI, code style, skill maintenance.
Auto-invocation: contributing and ci auto-invoke; new-recipe and manage-skills are manual-only.
Key principle: Reference CONTRIBUTING.md, tests/, .github/workflows/.
Which layer does this skill belong to? Skills should have a clear, non-overlapping scope. If it spans layers, split it.
Read existing skills in .claude/skills/ to ensure the new skill doesn't duplicate content. If there's overlap, update the existing skill instead.
Create .claude/skills/<skill-name>/SKILL.md with this structure:
---
name: <skill-name>
description: <Clear description of what the skill does and when to use it>
argument-hint: "[optional args]" # Only if the skill takes arguments
disable-model-invocation: true # Only for manual-trigger skills (Layer 4 actions)
---
# <Skill Title>
<Brief description of what this skill helps with>
## Step 1: Understand the request
<What to ask the user if not specified>
## Step 2: Reference existing code
<Which files to read for patterns — be specific with file paths>
## Step 3: Key concepts
<Core APIs, parameters, patterns>
## Step 4: Implementation
<Code examples following repo conventions>
## Step N: Add tests
<Testing guidance — smoke tests and unit tests>
Naming:
tinker-sdk, new-recipe, manage-skillsContent rules:
@chz.chz, explicit typing, etc.)/renderers skill" instead of re-explaining renderers)tests/recipes/ for smoke tests and *_test.py for unit testsFrontmatter rules:
description is required and must clearly state when to trigger the skilldisable-model-invocation: true only for action-oriented Layer 4 skillsargument-hint if the skill takes positional argumentsWhen auditing, check each skill for:
ls or grep to verify..claude/skills/
├── Layer 0: Fundamentals
│ ├── setup/ # Installation, API key, first run
│ ├── models/ # Model lineup, selection, families
│ ├── hyperparams/ # LR formulas, batch size, LoRA rank
│ └── logging/ # Training outputs, metrics, debugging
├── Layer 1: SDK
│ ├── tinker-sdk/ # ServiceClient, TrainingClient, SamplingClient, RestClient APIs
│ ├── tinker-types/ # Datum, ModelInput, TensorData, response types, error types
│ └── tinker-cli/ # tinker CLI: run/checkpoint management, download, publish
├── Layer 2: Primitives
│ ├── renderers/ # Renderer setup, TrainOnWhat, vision
│ ├── environments/ # Env, EnvGroupBuilder, custom RL envs
│ ├── weights/ # download, build_hf_model, publish
│ ├── completers/ # TokenCompleter, MessageCompleter
│ ├── checkpoints/ # save/load, CheckpointRecord, resume
│ ├── evals/ # Evaluators, Inspect AI
│ └── datasets/ # SupervisedDatasetBuilder, RLDatasetBuilder
├── Layer 3: Recipes
│ ├── sft/ # Supervised fine-tuning
│ ├── grpo/ # RL with verifiable rewards
│ ├── distillation/ # Knowledge distillation
│ ├── dpo/ # Direct Preference Optimization
│ ├── rlhf/ # RLHF pipeline
│ └── multiturn-rl/ # Multi-turn RL
└── Layer 4: Development
├── new-recipe/ # Scaffold new recipe
├── ci/ # Testing and CI
├── contributing/ # Dev setup and code style
└── manage-skills/ # This skill
When the codebase changes significantly (new modules, API changes, renamed files):
/manage-skills audit to check all skills还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Guide for checkpointing — saving, loading, and resuming training with CheckpointRecord. Use when the user asks about saving weights, resuming training, checkpoint management, or the checkpoint lifecycle.
中文简介正在准备中。当前显示原文说明。
Guide for testing conventions and CI pipelines — unit tests, integration smoke tests, pytest markers, and GitHub Actions workflows. Use when the user asks about testing, CI, running tests, or adding tests for a recipe.
中文简介正在准备中。当前显示原文说明。
Guide for using completers — TokenCompleter and MessageCompleter for text generation during RL rollouts and evaluation. Use when the user asks about generating text, completing messages, or using completers in RL environments.
中文简介正在准备中。当前显示原文说明。
Guide for contributing to the tinker-cookbook repo — development setup, code style, type checking, PR process, and design conventions. Use when the user asks about how to contribute, set up the dev environment, code style, or project conventions.
中文简介正在准备中。当前显示原文说明。
Guide for dataset construction — SupervisedDatasetBuilder, RLDatasetBuilder, ChatDatasetBuilder, and custom dataset creation from JSONL, HuggingFace, or conversation data. Use when the user asks about datasets, data loading, data preparation, or custom data formats.
中文简介正在准备中。当前显示原文说明。
Set up and run knowledge distillation (on-policy, off-policy, or multi-teacher) from a teacher model to a student model using the Tinker API. Use when the user wants to distill knowledge, compress models, or train a student from a teacher.
中文简介正在准备中。当前显示原文说明。