The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets for benchmarking, which presents a surface for indirect prompt injection.
Ingestion points: The skill loads datasets from external sources such as the Hugging Face Hub (e.g., the sharegpt dataset mentioned in references/flags.md) or from user-provided local paths.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to ignore or isolate embedded content within these datasets during the benchmarking process.
Capability inventory: The skill uses curl to query server health and the max benchmark CLI to generate load and write performance metrics to the local filesystem (e.g., results/latency.json).
Sanitization: No explicit sanitization or filtering of dataset content is described in the provided instructions.
Note: This ingestion is integral to the skill's primary function of model performance evaluation.
[COMMAND_EXECUTION]: The skill provides commands to execute local utilities for server management and benchmarking.
The curl command is used to verify the status of a local server at http://localhost:8000/v1/health and to retrieve served model names.
The max benchmark CLI tool is invoked to perform the actual load testing and performance measurement.
[EXTERNAL_DOWNLOADS]: The skill mentions the use of datasets hosted on the Hugging Face Hub (a well-known service) when selecting the sharegpt dataset, which involves downloading content during execution.

benchmark-model