The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including social media mentions, comments, and analytics exports. This content is used by the AI agent to generate draft replies and posts, creating a surface for indirect prompt injection attacks where malicious instructions embedded in incoming social media content could influence the agent's behavior.
Ingestion points: Data enters the system via scripts/ingest_snapshot.mjs (processing social snapshots) and is accessed by the agent through the engagement and posts bases in the Busabase provider (content/kelly-social-app/app/js/providers/busabase-provider.js).
Boundary markers: The skill employs a human-in-the-loop review process as the primary boundary. Instructions explicitly state that real publishing and replying happen only after human approval. A 'Quality Gate' (social-model.js) provides automated checks for brand voice and banned claims.
Capability inventory: The agent can draft posts, short-video scripts, and replies. The skill records these as 'intents' in the database. The companion UI allows for 'Publish' and 'Send Reply' actions, which are recorded as intent and executed out of band after approval.
Sanitization: The application uses escapeHtml for UI rendering and an automated evaluateGate function to filter content based on brand guidelines and specific banned claims (e.g., 'guaranteed', '100% secure').

kelly-social