The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the Cog installation script from the vendor's official domain at cog.run and the pget utility from the replicate/pget GitHub repository.
[REMOTE_CODE_EXECUTION]: Provides instructions to install the Cog CLI tool using a shell process substitution command: sh <(curl -fsSL https://cog.run/install.sh). This is the documented installation method from the tool's vendor.
[COMMAND_EXECUTION]: Demonstrates the use of subprocess.check_call and subprocess.run to execute the pget command-line tool for parallel downloading of model weights during the container setup phase.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of models that process untrusted user input via text prompts. 1. Ingestion points: The predict method in predict.py accepts a prompt string from external users. 2. Boundary markers: The documentation recommends implementing basic input validation and using the cog.Secret type for handling sensitive tokens. 3. Capability inventory: The model environment includes network access for weight fetching and the ability to execute code within the machine learning framework context. 4. Sanitization: The instructions focus on functional validation rather than defensive prompt sanitization, which is consistent with the skill's purpose as a development template.

build-models