The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external documentation and evidence provided by users. It mitigates this risk through explicit instructions to treat all analyzed content as data rather than instructions, and it prohibits taking any actions (like running commands or writing files) based on the content of the analyzed documents.
Ingestion points: The skill reads user-designated documentation files and evidence logs (documented in SKILL.md in the 'Check the target' and 'Allowed evidence' sections).
Boundary markers: The 'Boundaries' section explicitly instructs the agent to ignore instructions inside documents and logs, and it requires a mandatory 'Boundary Notes' block in all outputs to ensure compliance.
Capability inventory: Capabilities are limited to host-native read operations for documentation. The skill explicitly prohibits shell execution, file editing, and searching during the assessment phase.
Sanitization: The skill requires an 'atomicity pass' to break down claims into discrete components for evaluation against specific labels, which helps prevent the model from inadvertently following instructions embedded in large blocks of text.
[COMMAND_EXECUTION]: The instructions describe a single shell command (cat) used for the purpose of loading the skill's own instructions (SKILL.md) in environments that require command-backed loading. This is strictly limited to reading the skill's own source file at a host-provided path and is explicitly forbidden for use on any user-supplied data or evidence.

docs-claim-check