The Agent Skills Directory

[PROMPT_INJECTION]: The skill architecture handles untrusted external data (customer interactions, CRM records), creating a surface for indirect prompt injection. This is mitigated through a robust 'Governance Layer' and specific architectural guardrails:
Ingestion points: Data is ingested through the @personize/sdk via methods like memorizeBatch() in recipes/batch-pipeline.ts.
Boundary markers: The generation logic in reference/generate.md and recipes/generate-with-guardrails.ts uses structured context assembly with clear delimiters (e.g., ## Recipient Context) and explicitly calls smartGuidelines() to enforce organizational constraints over ingested data.
Capability inventory: Actions are limited to SDK operations and delivery via authorized third-party APIs (SendGrid, Slack, Twilio).
Sanitization: The skill provides automated post-generation checks (e.g., checkForHallucination in recipes/generate-with-guardrails.ts) and uses the SDK's evaluate: true parameter to validate output against safety and accuracy criteria.
[DATA_EXFILTRATION]: The skill requires the use of a PERSONIZE_SECRET_KEY and handles customer data. All network operations are directed to the vendor's official domain (personize.ai) or user-configured delivery endpoints (e.g., Slack webhooks, SendGrid). These operations are standard for the skill's primary purpose and do not represent unauthorized exfiltration.
[EXTERNAL_DOWNLOADS]: The skill references standard, well-known libraries such as @sendgrid/mail, @slack/web-api, and twilio. These dependencies are part of the intended integration surface for message delivery and originate from trusted registries.

personalization