The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands, including curl and Python script calls, using string interpolation of variables loaded from a JSON campaign plan.
Evidence: Step 3 explicitly defines shell commands such as curl -sS -o /dev/null -w "%{http_code}" "{landing_url}" and python "${CLAUDE_PLUGIN_ROOT}/scripts/crm-sync.py" --brand "{brand}" --plan ~/.claude-marketing/brands/{slug}/campaigns/{campaign_id}/plan.json.
Risk: If the campaign plan file (plan.json) is maliciously crafted or comes from an untrusted source, variables like {landing_url} or {brand} could contain shell metacharacters (e.g., backticks, semicolons) leading to arbitrary command execution on the host system.
[DATA_EXFILTRATION]: The skill performs outbound network requests to verify the availability of landing pages.
Evidence: In Step 3, the agent uses curl to reach out to the user-provided {landing_url} and retrieve HTTP status codes.
Risk: This capability allows the agent to interact with arbitrary external web servers. While used here for verification, it represents a standard network interaction surface that could be misused if the target URL is redirected or points to internal infrastructure.
[PROMPT_INJECTION]: The skill processes external data (campaign plans) that influences the execution flow and command arguments, creating an indirect injection surface.
Ingestion points: Campaign plan data is loaded from ~/.claude-marketing/brands/{slug}/campaigns/{campaign_id}/plan.json (Step 1).
Boundary markers: No specific delimiters or "ignore instructions" markers are used when interpolating the plan variables into shell commands.
Capability inventory: The skill has access to Bash (for curl and file operations) and can execute various Python scripts in the plugin root.
Sanitization: There is no explicit sanitization or validation logic described for the variables extracted from the campaign plan before they are used in command-line arguments.

launch-campaign