The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill accepts user-provided goal objectives that are stored in a local database and later interpolated into the agent's instructions during subsequent turns. This creates a surface where malicious instructions could be persisted and executed by the agent.\n
Ingestion points: User input provided via the /goal command is captured in SKILL.md (as $ARGUMENTS) and processed by scripts/claude_goal.py.\n
Boundary markers: The objective is delimited by <objective> tags in the prompt instructions generated by the Python script.\n
Capability inventory: The skill can execute local Python logic and modify a SQLite database. The agent itself possesses file system and command execution capabilities that could be targeted by a malicious objective.\n
Sanitization: The script enforces a 4000-character limit on objectives and the instructions explicitly warn the agent not to follow instructions within the objective that conflict with system or user messages.\n- [PERSISTENCE]: The skill implements a persistent state for goals and a mechanism to influence agent behavior across sessions.\n
State Storage: Goal data is persisted in a SQLite database located at ~/.claude/goal/goals.sqlite.\n
Execution Hook: The stop_hook function in scripts/claude_goal.py allows the skill to block the agent's termination signal when a goal is active, effectively maintaining the agent's focus until the goal is completed or paused.\n- [COMMAND_EXECUTION]: The skill relies on executing a local Python script to handle its core logic. User-provided arguments are passed to this script, though they are handled using shlex.split to mitigate shell injection risks during parsing.

goal