The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes multiple external data sources which could contain malicious instructions designed to influence agent behavior.
Ingestion points: The agent reads brand profiles from ~/.claude-marketing/brands/{slug}/profile.json, manifest files, and various guidelines (restrictions, messaging, voice-and-tone).
Boundary markers: There are no explicit delimiters or instructions to the agent to disregard instructions found within these files.
Capability inventory: The skill is capable of reading local files and executing python scripts via the command line.
Sanitization: No validation or sanitization of the external content is performed before interpolation into the agent's workflow.
[COMMAND_EXECUTION]: The skill specifies a shell command to be executed by the agent, incorporating a dynamic variable.
Evidence: The instruction python "${CLAUDE_PLUGIN_ROOT}/scripts/campaign-tracker.py" --brand {slug} --action list-campaigns in SKILL.md creates a potential command injection vector if the {slug} parameter is derived from untrusted user input without adequate validation.

funnel-architect