The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because its core functionality involves processing untrusted data from external sources (e.g., CRM records, support tickets, and meeting transcripts) and interpolating it into LLM prompts via the client.ai.prompt method.
Ingestion points include scripts in the recipes/ directory (e.g., batch-pipeline.ts, cold-outreach-sequence.ts) that fetch 'Contact' and 'Company' data.
Mitigation: The skill employs explicit boundary markers (e.g., ## Guidelines, ## Context) and provides a comprehensive 'GENERATE' guardrails framework in reference/generate.md to prevent hallucinations and ensure compliance.
[DATA_EXFILTRATION]: The skill facilitates the transmission of potentially sensitive generated content and entity metadata to external service providers and user-defined webhooks.
Scripts in recipes/ use the fetch API to send data to Slack webhooks (process.env.SLACK_WEBHOOK_URL) and generic webhook destinations.
The reference/wire.md and channels/ documentation detail how to connect Personize outputs to third-party APIs like SendGrid, Twilio, and Resend.
[COMMAND_EXECUTION]: The skill includes executable TypeScript 'recipes' designed to be run in the local environment. These scripts perform file system operations to maintain state.
Evidence: recipes/batch-pipeline.ts and recipes/health-check-message.ts use fs.readFileSync and fs.writeFileSync to manage JSON state files (e.g., .batch-pipeline-state.json) locally.
[EXTERNAL_DOWNLOADS]: The skill's functionality depends on the vendor-owned @personize/sdk and several well-known third-party Node.js packages for communication delivery.
Dependencies include @sendgrid/mail, twilio, resend, and @aws-sdk/client-ses.

solution-architect