The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs extremely imperative language ("ABSOLUTELY MUST", "NOT NEGOTIABLE", "YOU DO NOT HAVE A CHOICE") to override default agent behavior and mandate the use of the Skill tool. It explicitly states that these instructions "override default system prompt behavior."
[PROMPT_INJECTION]: The skill defines a "1% Rule" ("If you think there is even a 1% chance a skill might apply... you ABSOLUTELY MUST invoke the skill"). This policy forces the agent to load and potentially follow instructions from external files (skills) with minimal threshold for relevance, creating an entry point for indirect prompt injection.
Ingestion points: External skill files loaded via platform-specific tools (Skill, activate_skill).
Boundary markers: None. The agent is instructed to "Follow skill exactly."
Capability inventory: The skill references tools with high-impact capabilities including file modification (Write, Edit), shell execution (Bash), and subagent dispatch (Task).
Sanitization: No sanitization or validation logic is defined for the content of invoked skills.
[COMMAND_EXECUTION]: The mapping files (references/codex-tools.md, references/gemini-tools.md) provide direct instructions for executing shell commands for environment detection (e.g., git rev-parse) and platform-specific operations. While contextually appropriate for developer tools, these increase the skill's capability tier.

using-superpowers