The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data (e.g., shared documents, team channels, and messaging threads) to build document context. This creates a surface for indirect prompt injection where malicious instructions embedded in these sources could influence the agent's behavior.
Ingestion points: Stage 1: Initial Questions and Info Dumping in SKILL.md.
Boundary markers: Absent; there are no specific instructions to ignore or delimit embedded commands within ingested external files or chat logs.
Capability inventory: The skill uses create_file and str_replace in SKILL.md (Stage 2) to write to the user's workspace and invokes sub-agents in SKILL.md (Stage 3) for reader testing.
Sanitization: Absent; external content is interpolated into the workflow for brainstorming and drafting without sanitization or validation filters.
[DATA_EXFILTRATION]: The skill is designed to gather sensitive organizational context, including "team dynamics, past incidents, politics," and encourages the use of integrations (e.g., Slack, Google Drive, SharePoint) to pull this data into the agent's environment. While intended for document co-authoring, this involves processing high-value internal information.
[COMMAND_EXECUTION]: The skill instructs the agent to utilize filesystem tools such as create_file and str_replace. These capabilities allow the agent to create new document artifacts and modify existing files in the working directory based on user input and processed external data.

doc-coauthoring