The Agent Skills Directory

Indirect Prompt Injection (HIGH): The skill provides templates for agents (Support Triage, Sales Qualification, Marketing Coordinator) that ingest untrusted external data.
Ingestion points: SKILL.md identifies 'Ticket text', 'Inbound form', and 'Marketing brief' as primary inputs from external sources.
Capability inventory: Templates grant tools including Write, Edit, and Bash to agents processing this untrusted data. Specifically, the 'Ops Checklist Agent' is suggested to use Bash on system data, and the 'Marketing Coordinator' uses Write and Edit on campaign briefs.
Sanitization: No boundary markers, delimiters, or explicit sanitization instructions are provided in the skill to prevent the AI from obeying instructions embedded within that data.
External Script Execution (MEDIUM): The skill directs users to execute shell scripts that perform system-level operations without providing the script contents for audit.
Evidence: References to scripts/setup-vps.sh and scripts/deploy-agent.sh for configuring and deploying to a VPS.
Dependency Risks (LOW): The skill requires the installation of external packages.
Evidence: npm install @anthropic-ai/claude-agent-sdk express ws.
Status: Downgraded to LOW/INFO per [TRUST-SCOPE-RULE] because @anthropic-ai is a trusted organization and express/ws are standard industry packages.

agent-builder