The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from external sources (WhatsApp messages and clinical notes) and interpolates them into Large Language Model (LLM) prompts without sufficient sanitization.
Ingestion points: In src/ai/claude.service.ts, clinical notes and raw practitioner notes are fetched from the database and used as prompt context. In src/whatsapp/whatsapp.service.ts, incoming WhatsApp messages are processed.
Boundary markers: The prompts in ClaudeService (e.g., assistNoteGeneration and generateClinicalSummary) use basic string interpolation or simple quotes ("${rawNotes}") without robust delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has the ability to send WhatsApp messages, create appointments, generate AFIP invoices, and produce clinical risk assessments (detectRiskFactors).
Sanitization: There is no evidence of filtering or escaping clinical notes or user messages before they are processed by the AI.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions require cloning a repository from a personal GitHub account rather than a verified organization.
Evidence: git clone https://github.com/fahad-hamid/psique-workflow-clinic.git in the Installation section of SKILL.md.

sesion-mental-health-orchestration