The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of the clip-recipes package from PyPI. This is a vendor-specific resource associated with the author's stated tooling and is used to provide the underlying training and evaluation logic.
[COMMAND_EXECUTION]: The skill heavily utilizes command-line tools including torchrun and python -m to execute distributed training jobs, build datasets, and run evaluation suites. These commands involve significant file system access for writing checkpoints and managing data shards.
[PROMPT_INJECTION]: An indirect prompt injection surface is present as the skill processes external image-text data (e.g., captions) from local storage or remote S3 buckets. While the training process itself is a machine learning task, the agent's interaction with this untrusted content during data preparation and deduplication phases constitutes a risk.
Ingestion points: External text data is ingested via ShardBuilder and create_webdataset_loader (SKILL.md).
Boundary markers: The instructions do not define delimiters or specific 'ignore instructions' warnings for the data being processed.
Capability inventory: The skill has capabilities for shell command execution, network access (data streaming), and file system modification.
Sanitization: There are no explicit text sanitization or validation steps described for the captions being loaded into the pipeline.

clip-finetune-recipes