The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests potentially untrusted data from user-provided 'info dumps', external files, and links to shared documents.
Ingestion points: Context is gathered through direct text input, shared document links (e.g., Google Drive, SharePoint), and uploaded files within the Stage 1 workflow.
Boundary markers: The skill employs a 'Reader Testing' phase in Stage 3, which uses a fresh agent instance (or sub-agent) without conversation context to verify the document. This serves as a functional boundary to ensure the output is based only on the intended content.
Capability inventory: The agent utilizes standard file management tools such as create_file and str_replace to construct the document artifact, and sub-agent invocation to perform validation.
Sanitization: There are no explicit instructions to sanitize or escape the user-provided context before it is used to draft the document sections.
[SAFE]: The skill implements a defensive 'Reader Testing' strategy to identify ambiguities or contradictions in the generated document, which helps mitigate risks associated with processing complex external context.
[SAFE]: All external service references (e.g., Slack, Teams, Google Drive) are presented as platform-integrated connectors for legitimate context gathering and do not involve suspicious network requests or exfiltration patterns.

doc-coauthoring