The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a mechanism for persisting untrusted user data that is subsequently re-injected into the agent's active context via a companion hook.
Ingestion points: User conversation history and direct text provided through the /pin command logic in SKILL.md.
Boundary markers: The skill lacks specified delimiters, boundary markers, or 'ignore' instructions to isolate the pinned content from the agent's operational logic.
Capability inventory: The skill is configured with Bash, Read, Write, and Edit tools; it specifically uses shell commands for dynamic path derivation in the 'State File' section.
Sanitization: The instructions explicitly command the agent to 'Store verbatim' with 'no cleanup' and 'no reformulation', ensuring that any malicious prompts embedded in user decisions are preserved and re-presented to the model.

pin