The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs high-pressure directives to override the agent's default decision-making process. Instructions such as "YOU ABSOLUTELY MUST," "YOU DO NOT HAVE A CHOICE," and "This is not negotiable" are used to force the agent to invoke the Skill tool even when it might appear unnecessary. The documentation explicitly states that these skills "override default system prompt behavior."
[INDIRECT_PROMPT_INJECTION]: The skill establishes a systemic workflow that mandates loading external instructions from "skill" files before the agent provides any response or clarification to the user. This significantly increases the surface area for instructions from external files to influence or control agent behavior.
Ingestion points: The Skill (Claude Code), skill (Copilot), and activate_skill (Gemini) tools described in SKILL.md are used to pull in external instructional content.
Boundary markers: While the skill includes a priority list stating user instructions take precedence, it lacks explicit sanitization or "ignore embedded instructions" warnings for the content of the loaded skills themselves.
Capability inventory: The skill documentation references the use of powerful capabilities across platforms, including shell access (Bash), subagent dispatch (Task), and file system operations (Write, Edit), as seen in references/codex-tools.md and references/gemini-tools.md.
Sanitization: No validation or sanitization mechanism is described for the content fetched and interpreted by the skill-loading tools.

using-superpowers