The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script scripts/claude_goal.py to manage goal persistence, state transitions (active, paused, complete), and time/token accounting.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection by processing user-supplied objectives that are then injected into the agent's context as task instructions.
Ingestion points: User-provided goal descriptions are passed via the $ARGUMENTS variable in SKILL.md to the invoke command.
Boundary markers: The objective content is explicitly wrapped in <objective> tags within the instructions returned to the agent in scripts/claude_goal.py.
Capability inventory: The skill allows the agent to interact with a local SQLite database and can influence the agent's loop via a 'Stop' hook (decision: block) to ensure goals are pursued until completion.
Sanitization: While the skill validates the length of the objective, it does not sanitize content for prompt injection markers; however, it explicitly instructs the agent to treat the objective as 'task context' and to ignore instructions that conflict with higher-level constraints.

goal