The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of content from external and potentially untrusted sources to aid in document creation.
Ingestion points: The workflow in Stage 1 encourages the user to provide context by linking to shared documents, uploading files, or connecting to messaging integrations like Slack and Microsoft Teams.
Boundary markers: The instructions do not specify the use of delimiters or specific warnings to the agent to disregard embedded instructions within the ingested text, which could lead to the agent following malicious prompts hidden in the source material.
Capability inventory: The skill utilizes create_file and str_replace to write to the local file system and invokes sub-agents in Stage 3 to perform 'Reader Testing', creating a path where injected instructions could potentially influence file content or sub-agent behavior.
Sanitization: No explicit sanitization or validation logic is present for the data gathered from external integrations.
[COMMAND_EXECUTION]: The skill manages document drafting through standard file-system operations.
Evidence: The workflow explicitly directs the agent to use create_file for initial scaffolding and str_replace for iterative refinements of the document content. This behavior is aligned with the skill's primary objective of co-authoring documentation.
[EXTERNAL_DOWNLOADS]: The skill mentions fetching content from common productivity and messaging services.
Evidence: The instructions suggest using integrations for platforms such as Google Drive, SharePoint, Slack, and Microsoft Teams to gather project context. These references target well-known services and the data retrieval is context-dependent and initiated via user prompts.

doc-coauthoring