The Agent Skills Directory

[METADATA_POISONING]: The skill metadata contains conflicting information regarding its origin. It claims to be developed by 'ara.so', while the analysis context identifies the author as 'reason-machines', and the provided installation instructions point to a GitHub repository owned by an unrelated individual ('fahad-hamid').
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted clinical notes, patient history, and user queries by directly interpolating them into AI prompts.
Ingestion points: Clinical notes and patient history records processed in src/ai/claude.service.ts, and interactive user queries received in src/routes/ai-assistant/+page.svelte.
Boundary markers: Absent. The data is concatenated directly into natural language prompts without delimiters or specific instructions to disregard embedded commands.
Capability inventory: The skill has access to sensitive health records via src/prisma/prisma.service.ts and the ability to send messages via WhatsApp using src/whatsapp/whatsapp.service.ts.
Sanitization: Absent. No evidence of input validation, escaping, or structural delimiters to prevent embedded instructions from influencing agent behavior.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone source code from an external, non-vendor GitHub repository: https://github.com/fahad-hamid/psique-workflow-clinic.git.
[COMMAND_EXECUTION]: The installation process involves executing shell commands, including npm install, npx prisma migrate, and docker-compose up, which download and execute packages from the internet.

sesion-workflow-clinic