The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the local repository and external data sources retrieved via vendor tools, which can lead to indirect prompt injection.
Ingestion points: Inspects existing repository files, current versions, and external data sources or generative model outputs via the sandbase_run tool.
Boundary markers: The instructions do not explicitly mandate the use of delimiters or specific boundary markers when interpolating untrusted file content or external data into prompts.
Capability inventory: The skill is authorized to implement implementation changes (file writes), execute verification tests (shell command execution), and interact with external APIs (network access).
Sanitization: There are no instructions for sanitizing or filtering external content before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill uses vendor-specific tools to interact with external infrastructure for data discovery and task execution.
Evidence: Employs tools such as sandbase_discover, sandbase_inspect, and sandbase_run to fetch documentation and perform remote inference.
[COMMAND_EXECUTION]: The workflow requires the agent to execute tests and build checks to verify implementation success.
Evidence: The workflow instructs the agent to 'Run focused tests plus the relevant build or static checks' and 'Verify the real output with appropriate tests'.

behave-skill