The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses extremely forceful, imperative language to override default agent behavior. It includes directives such as 'YOU ABSOLUTELY MUST', 'NOT NEGOTIABLE', and 'YOU DO NOT HAVE A CHOICE', explicitly instructing the agent to ignore its own 'rationalization' and 'default system prompt' in favor of the 'Superpowers' skill workflow. It mandates the invocation of the Skill tool even if there is only a '1% chance' of applicability, attempting to bypass the agent's standard task assessment logic.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where external content from other skills and subagent prompt files is loaded and followed directly ('follow it directly', 'follow the instructions below exactly'). This creates an attack surface where a malicious skill file could influence the agent's behavior.
Ingestion points: Skill content loaded via the Skill tool and agent prompt files (.md) read from the filesystem.
Boundary markers: The skill suggests using XML tags like <agent-instructions> when framing messages for subagents, which provides some mitigation but does not eliminate the risk.
Capability inventory: The skill describes and facilitates the use of powerful tools including Bash, Task (subagent dispatch), Write, and Edit across multiple platforms (Claude Code, Copilot CLI, Gemini, Codex).
Sanitization: There is no evidence of input sanitization or validation of the content loaded from skill files or prompt templates before they are executed or passed to the agent.

using-superpowers