The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes extremely strong, imperative language designed to override the agent's core identity for role-playing. While these are intended for character consistency, terms like "MANDATORY", "CRITICAL", "NON-NEGOTIABLE", and instructions to "Ignore default Claude" behavior are patterns identical to those used in adversarial prompt injections to bypass system constraints.
[COMMAND_EXECUTION]: The skill includes a shell script, hooks/personality-hook.sh, that executes upon session start. It uses utilities like cat and jq to read configuration files and inject personality context into the agent's environment.
[DATA_EXFILTRATION]: The skill manages state by reading and writing to ~/.claude/personality.json. Accessing hidden configuration files in the user's home directory is a sensitive operation, though here it is limited to storing persona settings.
[PROMPT_INJECTION]: The personality system facilitates indirect prompt injection by dynamically loading instructions from external profile files into the agent's context.
Ingestion points: ~/.claude/personality.json and files within the profiles/ directory (e.g., profiles/mordecai.md).
Boundary markers: The hook script wraps injected content in <PERSONALITY_ACTIVE> tags to separate it from other instructions.
Capability inventory: The agent is instructed to read and update local configuration files and is explicitly told to re-read its own profile instructions if a version change is detected.
Sanitization: The shell hook uses jq -Rs . to escape content for JSON transmission, but it does not sanitize or validate the actual instruction content.
Evidence: The "VERSION CHECK" protocol in profiles/mordecai.md instructs the agent to "Read the ENTIRE profile file" and "Integrate behavioral changes" dynamically, allowing the agent's behavior to be modified via external file changes.

personality