[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79BB8HF9FBW7HVEM080Q2NK":3,"skill-same:01M79BB8HF9FBW7HVEM080Q2NK":65},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":31,"commitSha":32,"skillDirectory":33,"forksCount":13,"repositoryHtmlUrl":34,"repositoryHomepageUrl":6,"ownerHtmlUrl":35,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":36,"install":37,"resources":50},"01M79BB8HF9FBW7HVEM080Q2NK","convert-to-transformers",null,"Use this skill when converting custom PyTorch models to Hugging Face Transformers format. Helps with: (1) Creating PretrainedConfig and PreTrainedModel classes, (2) Writing ImageProcessor\u002FTokenizer, (3) Compatibility testing, (4) Hub upload preparation. Use when the user wants to make their model compatible with transformers library.","mryfmo\u002Fherness-self-improvement","mryfmo","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F11512262?v=4","01M71K2RDZFFQKYGMHMS5Y9PW1","01M79BAQKWWXZCC4B5C7M29WTS",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fconvert-to-transformers\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: convert-to-transformers\ndescription: \"Use this skill when converting custom PyTorch models to Hugging Face Transformers format. Helps with: (1) Creating PretrainedConfig and PreTrainedModel classes, (2) Writing ImageProcessor\u002FTokenizer, (3) Compatibility testing, (4) Hub upload preparation. Use when the user wants to make their model compatible with transformers library.\"\nmetadata:\n  scope: personal\n  owner: mryfmo\n  last-verified: 2026-07-21\n  freshness: 90d\n  provenance: repo=mryfmo\u002Fdotfiles; license=MIT\n  original-authorship: unknown\n  decision: D4\n  redistribution: prohibited\n---\n\n# Hugging Face Transformers Model Conversion\n\nConvert custom PyTorch models to Hugging Face Transformers format while maintaining exact compatibility with the original implementation.\n\n## Overview\n\nThis skill provides a systematic workflow for transformers conversion:\n\n- Extract hardcoded values into PretrainedConfig\n- Create PreTrainedModel wrapper\n- Build ImageProcessor\u002FTokenizer\n- Test equivalence thoroughly\n- Prepare for Hub upload\n\n**Important**: Use validation mode (parallel implementations) first to verify equivalence, then replace the original.\n\n## Conversion Workflow\n\n### Step 1: Analyze the Custom Model\n\nAsk the user to specify:\n\n- Path to the custom model implementation\n- Model type (vision, text, multimodal)\n- Task (classification, segmentation, generation, etc.)\n- Validation mode or replacement mode\n\nThen identify:\n\n- Model architecture and components\n- Input\u002Foutput formats\n- Key hyperparameters and hardcoded values\n- Pretrained weights location\n- Preprocessing pipeline\n- Custom layers or modules\n\n### Step 2: Create PretrainedConfig Class\n\n**Key principle**: Extract ALL hardcoded values from the model as configurable parameters.\n\nTemplate:\n\n```python\nfrom transformers import PretrainedConfig\nfrom typing import List, Optional\n\nclass {ModelName}Config(PretrainedConfig):\n    model_type = \"{model_name}\"\n\n    def __init__(\n        self,\n        # Core architecture parameters\n        hidden_dim: int = 128,\n        num_layers: int = 4,\n\n        # Input\u002Foutput parameters\n        image_size: int = 1024,\n        num_channels: int = 3,\n        num_labels: int = 1,\n\n        # Component-specific parameters (extract from original)\n        component_param: List[int] | None = None,\n\n        **kwargs,\n    ):\n        super().__init__(**kwargs)\n\n        self.hidden_dim = hidden_dim\n        self.num_layers = num_layers\n        self.image_size = image_size\n        self.num_channels = num_channels\n        self.num_labels = num_labels\n\n        # Use default if not specified\n        self.component_param = (\n            component_param if component_param is not None else [1, 2, 4]\n        )\n```\n\n**Critical**: Ensure default values match what the pretrained weights expect!\n\n### Step 3: Create PreTrainedModel Class\n\nTemplate:\n\n```python\nfrom transformers import PreTrainedModel\nfrom transformers.modeling_outputs import SemanticSegmenterOutput\n\nclass {ModelName}ForTask(PreTrainedModel):\n    config_class = {ModelName}Config\n\n    def __init__(self, config: {ModelName}Config):\n        super().__init__(config)\n        self.config = config\n\n        # Initialize layers using config parameters (no hardcoded values!)\n        self.encoder = Encoder(\n            hidden_dim=config.hidden_dim,\n            num_layers=config.num_layers,\n        )\n\n    def forward(\n        self,\n        pixel_values: torch.FloatTensor,\n        labels: Optional[torch.LongTensor] = None,\n        output_hidden_states: Optional[bool] = None,\n        return_dict: Optional[bool] = None,\n    ) -> Union[Tuple, SemanticSegmenterOutput]:\n        return_dict = return_dict if return_dict is not None else self.config.use_return_dict\n\n        # Forward pass\n        logits = self.encoder(pixel_values)\n\n        # Calculate loss if needed\n        loss = None\n        if labels is not None:\n            # Compute loss\n            pass\n\n        if not return_dict:\n            output = (logits,)\n            return ((loss,) + output) if loss is not None else output\n\n        return SemanticSegmenterOutput(\n            loss=loss,\n            logits=logits,\n            hidden_states=None,\n            attentions=None,\n        )\n```\n\n### Step 4: Create ImageProcessor\u002FTokenizer\n\n**For vision models** - Create ImageProcessor:\n\n```python\nfrom transformers import BaseImageProcessor\n\nclass {ModelName}ImageProcessor(BaseImageProcessor):\n    model_input_names = [\"pixel_values\"]\n\n    def __init__(\n        self,\n        size: int = 1024,\n        resample: str = \"bilinear\",\n        do_normalize: bool = True,\n        image_mean: List[float] | None = None,\n        image_std: List[float] | None = None,\n        **kwargs,\n    ):\n        super().__init__(**kwargs)\n        self.size = size\n        self.resample = resample\n        self.do_normalize = do_normalize\n        self.image_mean = image_mean if image_mean is not None else [0.485, 0.456, 0.406]\n        self.image_std = image_std if image_std is not None else [0.229, 0.224, 0.225]\n\n    def preprocess(\n        self,\n        images: ImageInput,\n        return_tensors: Optional[Union[str, TensorType]] = None,\n        **kwargs,\n    ) -> BatchFeature:\n        # Implement preprocessing matching original\n        # Return BatchFeature with pixel_values\n```\n\n**For text models** - Create tokenizer configuration.\n\n### Step 5: Create Compatibility Tests\n\n**Critical**: Always use the SAME preprocessed tensor for both models when comparing outputs.\n\n```python\nimport pytest\nimport torch\n\ndef test_preprocessing_matches():\n    \"\"\"Test that preprocessing is equivalent.\"\"\"\n    old_tensor = old_preprocessing(image)\n    new_tensor = processor(image, return_tensors=\"pt\")[\"pixel_values\"][0]\n    assert torch.allclose(old_tensor, new_tensor, atol=1e-6)\n\ndef test_single_image_output_matches():\n    \"\"\"Test that model outputs match.\"\"\"\n    # Load models\n    old_model = OldModel()\n    new_model = NewModel(NewConfig())\n    new_model.load_state_dict(old_model.state_dict())\n\n    # Prepare SAME input\n    preprocessed = preprocess(image)\n\n    with torch.no_grad():\n        old_output = old_model(preprocessed)\n        new_output = new_model(pixel_values=preprocessed)\n\n    # Use 0.5% tolerance for numerical differences\n    assert torch.allclose(old_output, new_output.logits, atol=5e-3, rtol=1e-2)\n\ndef test_batch_output_matches():\n    \"\"\"Test batch processing.\"\"\"\n    # Test with batch of images\n\ndef test_state_dict_compatible():\n    \"\"\"Test that weights can be loaded.\"\"\"\n    new_model.load_state_dict(old_model.state_dict())\n```\n\n### Step 6: Create MODEL_CARD.md\n\nGenerate comprehensive model card following Hugging Face standards. Include:\n\n- Model description\n- Usage examples\n- Training details\n- Evaluation metrics\n- Citation information\n\nSee Hugging Face model card documentation for template.\n\n### Step 7: Create Hub Push Script\n\n**Critical**: Register classes with `register_for_auto_class()` before pushing.\n\n```python\n#!\u002Fusr\u002Fbin\u002Fenv python3\nfrom huggingface_hub import HfApi\nfrom {module}.transformers import {ModelName}Config, {ModelName}ForTask, {ModelName}ImageProcessor\n\ndef main():\n    # CRITICAL: Register for Auto* support\n    {ModelName}Config.register_for_auto_class()\n    {ModelName}ForTask.register_for_auto_class(\"AutoModel\")\n    {ModelName}ImageProcessor.register_for_auto_class(\"AutoImageProcessor\")\n\n    # Load original model\n    original_model = OriginalModel()\n\n    # Create transformers-compatible model\n    config = {ModelName}Config()\n    model = {ModelName}ForTask(config)\n    model.load_state_dict(original_model.state_dict())\n\n    # Save and push\n    model.save_pretrained(local_dir)\n    config.save_pretrained(local_dir)\n    processor = {ModelName}ImageProcessor()\n    processor.save_pretrained(local_dir)\n\n    api = HfApi(token=token)\n    api.create_repo(repo_id=repo_id, exist_ok=True)\n    api.upload_folder(repo_id=repo_id, folder_path=local_dir)\n```\n\n## Implementation Strategy\n\n### Validation Mode (Recommended First)\n\nCreate parallel implementations:\n\n```\n{project}\u002F\n├── {module}\u002F\n│   ├── original_model.py           # Existing\n│   └── transformers\u002F               # NEW - for validation\n│       ├── __init__.py\n│       ├── configuration_{model}.py\n│       ├── modeling_{model}.py\n│       └── processing_{model}.py\n└── tests\u002F\n    └── test_transformers_compatibility.py\n```\n\nWorkflow:\n\n1. Create transformers\u002F package alongside original\n2. Run compatibility tests\n3. Debug any discrepancies\n4. Once tests pass → proceed to replacement\n\n### Replacement Mode (After Validation)\n\nOnce equivalence is verified:\n\n```\n{project}\u002F\n└── {module}\u002F\n    ├── configuration_{model}.py    # Replaces original\n    ├── modeling_{model}.py\n    └── processing_{model}.py\n```\n\nWorkflow:\n\n1. Remove or archive original implementation\n2. Move transformers\u002F\\* files up one level\n3. Update all imports\n4. Update tests\n5. Update documentation\n\n## Common Issues\n\nFor detailed troubleshooting, see [references\u002Fcommon-pitfalls.md](references\u002Fcommon-pitfalls.md).\n\nQuick reference:\n\n- **Hardcoded values**: Extract to config with matching defaults\n- **Preprocessing mismatches**: Use exact same pipeline and parameters\n- **State dict keys**: Keep layer names matching original\n- **Test tolerance**: Use 0.5% tolerance (5e-3) for numerical differences\n- **Device handling**: Use `self.device` from PreTrainedModel\n- **Image size order**: `post_process_semantic_segmentation()` expects `(width, height)`\n\n## Debugging Strategy\n\nIf outputs don't match:\n\n1. Check preprocessing produces identical tensors\n2. Check state dict loaded correctly\n3. Step-by-step comparison of each layer\n4. Check device placement\n5. Check determinism with torch.manual_seed\n\nSee [references\u002Fcommon-pitfalls.md](references\u002Fcommon-pitfalls.md) for detailed debugging steps.\n\n## Project Learnings\n\nAfter completing a conversion, add learnings to [references\u002Flearnings.md](references\u002Flearnings.md).\n\nThis accumulates knowledge from each project to avoid repeating mistakes.\n\n## Checklist\n\n- [ ] Step 1: Model analyzed\n  - [ ] Architecture identified\n  - [ ] Hyperparameters extracted\n  - [ ] Preprocessing understood\n- [ ] Step 2: Config class created\n  - [ ] All hardcoded values moved to config\n  - [ ] Defaults match pretrained weights\n  - [ ] Type hints added\n- [ ] Step 3: Model class created\n  - [ ] Inherits from PreTrainedModel\n  - [ ] Uses only config parameters\n  - [ ] Forward signature matches conventions\n- [ ] Step 4: Processor created\n  - [ ] Preprocessing matches original\n  - [ ] Returns appropriate format\n- [ ] Step 5: Tests created\n  - [ ] Preprocessing test passes\n  - [ ] Single image test passes\n  - [ ] Batch test passes\n  - [ ] State dict test passes\n- [ ] Step 6: MODEL_CARD.md created\n  - [ ] All sections filled\n  - [ ] Usage examples tested\n  - [ ] Citations included\n- [ ] Step 7: Push script created\n  - [ ] register_for_auto_class() called\n  - [ ] Script tested locally\n  - [ ] Successfully pushed to Hub\n- [ ] Cleanup (if doing replacement)\n  - [ ] Original implementation removed\u002Farchived\n  - [ ] Imports updated\n  - [ ] Documentation updated\n","# Hugging Face Transformers Model Conversion\n\nConvert custom PyTorch models to Hugging Face Transformers format while maintaining exact compatibility with the original implementation.\n\n## Overview\n\nThis skill provides a systematic workflow for transformers conversion:\n\n- Extract hardcoded values into PretrainedConfig\n- Create PreTrainedModel wrapper\n- Build ImageProcessor\u002FTokenizer\n- Test equivalence thoroughly\n- Prepare for Hub upload\n\n**Important**: Use validation mode (parallel implementations) first to verify equivalence, then replace the original.\n\n## Conversion Workflow\n\n### Step 1: Analyze the Custom Model\n\nAsk the user to specify:\n\n- Path to the custom model implementation\n- Model type (vision, text, multimodal)\n- Task (classification, segmentation, generation, etc.)\n- Validation mode or replacement mode\n\nThen identify:\n\n- Model architecture and components\n- Input\u002Foutput formats\n- Key hyperparameters and hardcoded values\n- Pretrained weights location\n- Preprocessing pipeline\n- Custom layers or modules\n\n### Step 2: Create PretrainedConfig Class\n\n**Key principle**: Extract ALL hardcoded values from the model as configurable parameters.\n\nTemplate:\n\n```python\nfrom transformers import PretrainedConfig\nfrom typing import List, Optional\n\nclass {ModelName}Config(PretrainedConfig):\n    model_type = \"{model_name}\"\n\n    def __init__(\n        self,\n        # Core architecture parameters\n        hidden_dim: int = 128,\n        num_layers: int = 4,\n\n        # Input\u002Foutput parameters\n        image_size: int = 1024,\n        num_channels: int = 3,\n        num_labels: int = 1,\n\n        # Component-specific parameters (extract from original)\n        component_param: List[int] | None = None,\n\n        **kwargs,\n    ):\n        super().__init__(**kwargs)\n\n        self.hidden_dim = hidden_dim\n        self.num_layers = num_layers\n        self.image_size = image_size\n        self.num_channels = num_channels\n        self.num_labels = num_labels\n\n        # Use default if not specified\n        self.component_param = (\n            component_param if component_param is not None else [1, 2, 4]\n        )\n```\n\n**Critical**: Ensure default values match what the pretrained weights expect!\n\n### Step 3: Create PreTrainedModel Class\n\nTemplate:\n\n```python\nfrom transformers import PreTrainedModel\nfrom transformers.modeling_outputs import SemanticSegmenterOutput\n\nclass {ModelName}ForTask(PreTrainedModel):\n    config_class = {ModelName}Config\n\n    def __init__(self, config: {ModelName}Config):\n        super().__init__(config)\n        self.config = config\n\n        # Initialize layers using config parameters (no hardcoded values!)\n        self.encoder = Encoder(\n            hidden_dim=config.hidden_dim,\n            num_layers=config.num_layers,\n        )\n\n    def forward(\n        self,\n        pixel_values: torch.FloatTensor,\n        labels: Optional[torch.LongTensor] = None,\n        output_hidden_states: Optional[bool] = None,\n        return_dict: Optional[bool] = None,\n    ) -> Union[Tuple, SemanticSegmenterOutput]:\n        return_dict = return_dict if return_dict is not None else self.config.use_return_dict\n\n        # Forward pass\n        logits = self.encoder(pixel_values)\n\n        # Calculate loss if needed\n        loss = None\n        if labels is not None:\n            # Compute loss\n            pass\n\n        if not return_dict:\n            output = (logits,)\n            return ((loss,) + output) if loss is not None else output\n\n        return SemanticSegmenterOutput(\n            loss=loss,\n            logits=logits,\n            hidden_states=None,\n            attentions=None,\n        )\n```\n\n### Step 4: Create ImageProcessor\u002FTokenizer\n\n**For vision models** - Create ImageProcessor:\n\n```python\nfrom transformers import BaseImageProcessor\n\nclass {ModelName}ImageProcessor(BaseImageProcessor):\n    model_input_names = [\"pixel_values\"]\n\n    def __init__(\n        self,\n        size: int = 1024,\n        resample: str = \"bilinear\",\n        do_normalize: bool = True,\n        image_mean: List[float] | None = None,\n        image_std: List[float] | None = None,\n        **kwargs,\n    ):\n        super().__init__(**kwargs)\n        self.size = size\n        self.resample = resample\n        self.do_normalize = do_normalize\n        self.image_mean = image_mean if image_mean is not None else [0.485, 0.456, 0.406]\n        self.image_std = image_std if image_std is not None else [0.229, 0.224, 0.225]\n\n    def preprocess(\n        self,\n        images: ImageInput,\n        return_tensors: Optional[Union[str, TensorType]] = None,\n        **kwargs,\n    ) -> BatchFeature:\n        # Implement preprocessing matching original\n        # Return BatchFeature with pixel_values\n```\n\n**For text models** - Create tokenizer configuration.\n\n### Step 5: Create Compatibility Tests\n\n**Critical**: Always use the SAME preprocessed tensor for both models when comparing outputs.\n\n```python\nimport pytest\nimport torch\n\ndef test_preprocessing_matches():\n    \"\"\"Test that preprocessing is equivalent.\"\"\"\n    old_tensor = old_preprocessing(image)\n    new_tensor = processor(image, return_tensors=\"pt\")[\"pixel_values\"][0]\n    assert torch.allclose(old_tensor, new_tensor, atol=1e-6)\n\ndef test_single_image_output_matches():\n    \"\"\"Test that model outputs match.\"\"\"\n    # Load models\n    old_model = OldModel()\n    new_model = NewModel(NewConfig())\n    new_model.load_state_dict(old_model.state_dict())\n\n    # Prepare SAME input\n    preprocessed = preprocess(image)\n\n    with torch.no_grad():\n        old_output = old_model(preprocessed)\n        new_output = new_model(pixel_values=preprocessed)\n\n    # Use 0.5% tolerance for numerical differences\n    assert torch.allclose(old_output, new_output.logits, atol=5e-3, rtol=1e-2)\n\ndef test_batch_output_matches():\n    \"\"\"Test batch processing.\"\"\"\n    # Test with batch of images\n\ndef test_state_dict_compatible():\n    \"\"\"Test that weights can be loaded.\"\"\"\n    new_model.load_state_dict(old_model.state_dict())\n```\n\n### Step 6: Create MODEL_CARD.md\n\nGenerate comprehensive model card following Hugging Face standards. Include:\n\n- Model description\n- Usage examples\n- Training details\n- Evaluation metrics\n- Citation information\n\nSee Hugging Face model card documentation for template.\n\n### Step 7: Create Hub Push Script\n\n**Critical**: Register classes with `register_for_auto_class()` before pushing.\n\n```python\n#!\u002Fusr\u002Fbin\u002Fenv python3\nfrom huggingface_hub import HfApi\nfrom {module}.transformers import {ModelName}Config, {ModelName}ForTask, {ModelName}ImageProcessor\n\ndef main():\n    # CRITICAL: Register for Auto* support\n    {ModelName}Config.register_for_auto_class()\n    {ModelName}ForTask.register_for_auto_class(\"AutoModel\")\n    {ModelName}ImageProcessor.register_for_auto_class(\"AutoImageProcessor\")\n\n    # Load original model\n    original_model = OriginalModel()\n\n    # Create transformers-compatible model\n    config = {ModelName}Config()\n    model = {ModelName}ForTask(config)\n    model.load_state_dict(original_model.state_dict())\n\n    # Save and push\n    model.save_pretrained(local_dir)\n    config.save_pretrained(local_dir)\n    processor = {ModelName}ImageProcessor()\n    processor.save_pretrained(local_dir)\n\n    api = HfApi(token=token)\n    api.create_repo(repo_id=repo_id, exist_ok=True)\n    api.upload_folder(repo_id=repo_id, folder_path=local_dir)\n```\n\n## Implementation Strategy\n\n### Validation Mode (Recommended First)\n\nCreate parallel implementations:\n\n```\n{project}\u002F\n├── {module}\u002F\n│   ├── original_model.py           # Existing\n│   └── transformers\u002F               # NEW - for validation\n│       ├── __init__.py\n│       ├── configuration_{model}.py\n│       ├── modeling_{model}.py\n│       └── processing_{model}.py\n└── tests\u002F\n    └── test_transformers_compatibility.py\n```\n\nWorkflow:\n\n1. Create transformers\u002F package alongside original\n2. Run compatibility tests\n3. Debug any discrepancies\n4. Once tests pass → proceed to replacement\n\n### Replacement Mode (After Validation)\n\nOnce equivalence is verified:\n\n```\n{project}\u002F\n└── {module}\u002F\n    ├── configuration_{model}.py    # Replaces original\n    ├── modeling_{model}.py\n    └── processing_{model}.py\n```\n\nWorkflow:\n\n1. Remove or archive original implementation\n2. Move transformers\u002F\\* files up one level\n3. Update all imports\n4. Update tests\n5. Update documentation\n\n## Common Issues\n\nFor detailed troubleshooting, see [references\u002Fcommon-pitfalls.md](references\u002Fcommon-pitfalls.md).\n\nQuick reference:\n\n- **Hardcoded values**: Extract to config with matching defaults\n- **Preprocessing mismatches**: Use exact same pipeline and parameters\n- **State dict keys**: Keep layer names matching original\n- **Test tolerance**: Use 0.5% tolerance (5e-3) for numerical differences\n- **Device handling**: Use `self.device` from PreTrainedModel\n- **Image size order**: `post_process_semantic_segmentation()` expects `(width, height)`\n\n## Debugging Strategy\n\nIf outputs don't match:\n\n1. Check preprocessing produces identical tensors\n2. Check state dict loaded correctly\n3. Step-by-step comparison of each layer\n4. Check device placement\n5. Check determinism with torch.manual_seed\n\nSee [references\u002Fcommon-pitfalls.md](references\u002Fcommon-pitfalls.md) for detailed debugging steps.\n\n## Project Learnings\n\nAfter completing a conversion, add learnings to [references\u002Flearnings.md](references\u002Flearnings.md).\n\nThis accumulates knowledge from each project to avoid repeating mistakes.\n\n## Checklist\n\n- [ ] Step 1: Model analyzed\n  - [ ] Architecture identified\n  - [ ] Hyperparameters extracted\n  - [ ] Preprocessing understood\n- [ ] Step 2: Config class created\n  - [ ] All hardcoded values moved to config\n  - [ ] Defaults match pretrained weights\n  - [ ] Type hints added\n- [ ] Step 3: Model class created\n  - [ ] Inherits from PreTrainedModel\n  - [ ] Uses only config parameters\n  - [ ] Forward signature matches conventions\n- [ ] Step 4: Processor created\n  - [ ] Preprocessing matches original\n  - [ ] Returns appropriate format\n- [ ] Step 5: Tests created\n  - [ ] Preprocessing test passes\n  - [ ] Single image test passes\n  - [ ] Batch test passes\n  - [ ] State dict test passes\n- [ ] Step 6: MODEL_CARD.md created\n  - [ ] All sections filled\n  - [ ] Usage examples tested\n  - [ ] Citations included\n- [ ] Step 7: Push script created\n  - [ ] register_for_auto_class() called\n  - [ ] Script tested locally\n  - [ ] Successfully pushed to Hub\n- [ ] Cleanup (if doing replacement)\n  - [ ] Original implementation removed\u002Farchived\n  - [ ] Imports updated\n  - [ ] Documentation updated\n",{"name":5,"metadata":23,"description":7},{"owner":9,"scope":24,"decision":25,"freshness":26,"provenance":27,"last-verified":28,"redistribution":29,"original-authorship":30},"personal","D4","90d","repo=mryfmo\u002Fdotfiles; license=MIT",1784592000000,"prohibited","unknown","main","8b4cc13bc73f8d52e6aa54d860c9d345b31d4032",".claude\u002Fskills\u002Fconvert-to-transformers","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement","https:\u002F\u002Fgithub.com\u002Fmryfmo","2026-07-21 09:19:35",{"source":8,"skillName":5,"commands":38,"prompt":48,"zipAvailable":49},[39,42,45],{"tool":40,"command":41},"npx","npx skills add mryfmo\u002Fherness-self-improvement --skill convert-to-transformers",{"tool":43,"command":44},"bunx","bunx skills add mryfmo\u002Fherness-self-improvement --skill convert-to-transformers",{"tool":46,"command":47},"pnpm","pnpm dlx skills add mryfmo\u002Fherness-self-improvement --skill convert-to-transformers","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fconvert-to-transformers\u002FSKILL.md",true,[51,57,61],{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"eval\u002FNOTES.md","REFERENCE","参考文档",170,"https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fconvert-to-transformers\u002Feval\u002FNOTES.md",{"path":58,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":59,"sourceHtmlUrl":60},"references\u002Fcommon-pitfalls.md",5689,"https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fconvert-to-transformers\u002Freferences\u002Fcommon-pitfalls.md",{"path":62,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":63,"sourceHtmlUrl":64},"references\u002Flearnings.md",3701,"https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fconvert-to-transformers\u002Freferences\u002Flearnings.md",[66,73,80,87,94,101],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":18},"01M79BBVDZ6CHDG1Z9GMV5EJA6","shdoc-shell-docs","Write and review shellscript documentation with shdoc annotations. Use when Codex creates, edits, or reviews `.sh` files or shell executables and should add, repair, or normalize `@file`, `@brief`, `@description`, `@arg`, `@option`, and `@example` comments to match shdoc conventions.","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fshdoc-shell-docs\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":18},"01M79BBRPXW9ZSYKP02N6K7M4G","python-uv-workflow","Apply Python development policy using uv-first execution, test-first behavior validation, and pre-commit quality gates. Use when implementing or refactoring Python code.","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fpython-uv-workflow\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":18},"01M79BBNZB6B9R8KNQV5V4SCJM","humanizer-ja","AIくさい日本語を、人が書いたように自然でこなれた日本語へ書き換えるスキル。ブログ記事、メモ、メール、文書、SNS投稿、チャット返信などの日本語を書いたり推敲したりするときに使う。AIっぽい言い回しを消したいとき、日本語を自然にしたいとき、言い換えたいとき、硬さを抜きたいとき、トーンを整えたいとき、人間らしい声や温度を足したいときに使う。","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fhumanizer-ja\u002FSKILL.md",[],[],{"skillNo":88,"name":89,"nameZh":6,"description":90,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":18},"01M79BBJPWD4SH8EFC0CAWH64H","gh-first-workflow","Enforce gh-first GitHub investigation, pull request maintenance, and Conventional Commit output rules. Use when investigating GitHub issues or pull requests, creating or updating pull requests, summarizing investigation results, or preparing commit messages.","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fgh-first-workflow\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":98,"categories":99,"keywords":100,"publishedAt":18},"01M79BBF1KAC97FM9T2PDB4Z4B","gh-comment-attach-files","Attach local files to a GitHub issue or pull request comment via Playwright CLI and return the hosted attachment URLs without submitting the comment. Use when you need GitHub-hosted image or document URLs on github.com or GitHub Enterprise Server.","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fgh-comment-attach-files\u002FSKILL.md",[],[],{"skillNo":102,"name":103,"nameZh":6,"description":104,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":105,"categories":106,"keywords":107,"publishedAt":18},"01M79BB3KY0WXQDD95NJ80JZ4V","agmsg","Cross-agent messaging via SQLite. Send messages between Claude Code, Codex, Gemini CLI, and other agents. No daemon, no network, no dependencies beyond bash and sqlite3.","https:\u002F\u002Fgithub.com\u002Fmryfmo\u002Fherness-self-improvement\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fagmsg\u002FSKILL.md",[],[]]