The Agent Skills Directory

[SAFE]: The skill serves as a protocol for using the AppFollow MCP tools. It does not include any hidden commands, obfuscated code, or unauthorized network requests. All interactions are conducted through the defined MCP server interface using tools mapped in the documentation.
[DATA_EXPOSURE]: The skill includes a feedback mechanism (submit_feedback) that allows users to send session transcripts and tool logs to AppFollow. This process is protected by multiple safeguards: the agent must obtain explicit consent for each type of data shared, and instructions explicitly warn against including any secrets or credentials. This is a legitimate vendor feature with built-in privacy controls.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (app reviews) and possesses write capabilities (such as replying to reviews). It successfully mitigates indirect prompt injection risks by enforcing a human-in-the-loop 'preview → approval → confirm' process for all write tools. This prevents the agent from automatically executing instructions that might be embedded in review content.
Ingestion points: External data is ingested through the get_reviews and get_reviews_ai_summary tools.
Boundary markers: The mandatory approval ritual acts as a verification boundary between ingested data and agent actions.
Capability inventory: Write capabilities include reply_to_review, edit_keywords, and add_app tools.
Sanitization: The instructions provide clear warnings to both the agent and the user regarding the exclusion of sensitive information from free-text fields.

using-appfollow