The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted content from the repository and user inputs. \n
Ingestion points: Inspects available files, repository instructions, current versions, and user changes as defined in the Workflow and Quality Gates (SKILL.md). \n
Boundary markers: The skill requires restating outcomes and constraints and performing source comparisons, though it lacks explicit technical delimiters for processed data. \n
Capability inventory: Utilizes the sandbase_run tool to execute discovered capabilities and sandbase_run_get for polling results. \n
Sanitization: Mandates sending only the minimum necessary data to external tools and requires obtaining explicit user confirmation before any paid or multi-call execution. \n- [EXTERNAL_DOWNLOADS]: Interacts with vendor-owned infrastructure via sandbase_discover and sandbase_inspect to retrieve capability schemas and external documentation. \n- [COMMAND_EXECUTION]: The sandbase_run tool executes tasks discovered at runtime. The skill includes protective measures by requiring the agent to display endpoints, arguments, and pricing to the user for approval before execution.

google-cloud-waf-sustainability