The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script using string interpolation for arguments in Step 4: python "${CLAUDE_PLUGIN_ROOT}/scripts/narrative-mapper.py" --brand {slug} --action map-landscape --data '{positions_json}'.
Evidence: This pattern is susceptible to command injection if the {positions_json} variable, which is populated from external web content, contains shell metacharacters.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from competitor websites, advertising copy, and social media (Step 3).
Ingestion points: External URLs provided by the user or identified by the agent for competitor analysis.
Boundary markers: No explicit instruction to use delimiters or ignore embedded instructions in the ingested data is present.
Capability inventory: The skill has the capability to execute shell commands (via the Python script) and read local filesystem data.
Sanitization: No sanitization or validation steps for the external content are defined before it is processed or used in script arguments.
[DATA_EXFILTRATION]: The skill accesses sensitive-looking local files such as ~/.claude-marketing/brands/{slug}/profile.json and ~/.claude-marketing/brands/{slug}/guidelines/_manifest.json.
Evidence: While these are application-specific paths, the skill identifies and loads these files in Step 1 to extract brand positioning and guardrails, which are then processed alongside external data.

narrative-landscape