The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs aggressive, non-negotiable language to override the agent's internal decision-making process. It explicitly states that "Superpowers skills override default system prompt behavior" and uses forceful imperatives like "YOU ABSOLUTELY MUST invoke the skill," "This is not negotiable," and "This is not optional." This pattern is designed to force the AI to adhere to the skill's framework over its pre-trained safety or operational guidelines.
[DYNAMIC_EXECUTION]: The skill defines a mechanism for the dynamic loading and execution of instructions from external SKILL.md files. These files are sourced from various computed paths, including project-level, user-global (~/.claude/skills/, ~/.codex/skills/, etc.), and platform-specific directories. This behavior allows for the runtime modification of the agent's behavior based on files found in the environment.
[INDIRECT_PROMPT_INJECTION]: The skill creates a vulnerability surface by requiring the agent to search for and invoke skills for almost any task (even with a "1% chance" of relevance). This "invoke-first" rule, combined with the instruction to "Follow skill exactly," increases the risk that a malicious skill file placed in a scanned directory could be executed without sufficient verification, potentially leading to harmful actions.
Ingestion points: Discovery and loading of SKILL.md files from multiple directory levels across different runtimes (Claude Code, Codex, Copilot, Gemini, Pi, Antigravity).
Boundary markers: The instructions use specific XML-like tags (<EXTREMELY-IMPORTANT>, <SUBAGENT-STOP>) to denote priority and scope, but these do not serve as safety boundaries for external content.
Capability inventory: The platform reference files map actions to high-capability tools including shell access (Bash, run_command, shell, run_shell_command) and network operations (WebFetch, WebSearch, curl).
Sanitization: There is no evidence of content validation or sanitization for the loaded skill instructions before they are followed by the agent.
[COMMAND_EXECUTION]: The reference documentation explicitly maps abstract actions to powerful command-line tools across various platforms. While the mapping itself is for documentation, the skill's purpose is to guide the agent in using these tools to execute the instructions found in the dynamically loaded skill files.

using-superpowers