The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses files within a hidden application configuration directory in the user's home folder, which may contain sensitive configuration data for the agent environment.
Evidence: Reads and writes to ~/.claude/ad-profiles/config.json and ~/.claude/ad-profiles/[brand-slug].md to maintain persistent state across sessions.
Evidence: Accesses ~/.claude/skills/brand-voice/ to retrieve voice profile data.
[INDIRECT_PROMPT_INJECTION]: The skill accepts untrusted data (such as user reviews or direct messages) and stores it in profile files that are re-read into the agent's context during subsequent executions.
Ingestion points: Processes user-provided text in the "interview" phase (Step 0) and reads stored profile data and learnings.md in Step 1 and Step 2.
Boundary markers: Absent; the skill relies on standard markdown headers without explicit delimiters to separate stored data from instructions.
Capability inventory: Includes the ability to read and write files in designated directories.
Sanitization: No sanitization or validation logic is specified for the raw language or proof inputs provided by users before they are saved to the profile.

sales-page