Claude Skills by feiyang-k

Use this skill when you want to create a large-scale dataset with region-level recognition and text descriptions for panoptic visual understanding. Avoid it when you do not need region-level understanding or panoptic annotations.

Use this skill when you need part-level segmentation data for training VLMs to understand object parts and fine-grained visual components. Avoid it when object-level segmentation is sufficient.

Use this skill when you want the simplest possible automated augmentation: randomly select one transform and one magnitude per image, no tuning needed. Avoid it when you prefer tuned augmentation policies.

Use this skill when you want to generate diverse instruction data by prompting an LLM with creative seed constraints for novel instruction types. Avoid it when standard self-instruct provides sufficient diversity.

Use this skill when you want to generate video instruction data by prompting GPT-3.5 with video descriptions and annotations for video understanding VLM training. Avoid it when you do not work with video or have sufficient video instruction data.

Use this skill when you want to generate multimodal instruction-following data by prompting a text-only LLM with image captions and bounding boxes. Avoid it when you have abundant human-annotated instruction data or need the LLM to see actual images during data generation.

Use this skill when you want to augment text training data with visual tokens ('vokens') retrieved from an image database to ground language in vision. Avoid it when you do not need visual grounding of text data.

Use this skill when you want to progressively increase instruction complexity through evolution prompts (deepening, widening, adding constraints). Avoid it when you need simple instructions or cannot afford iterative evolution.