The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses imperative and forceful language (e.g., 'ABSOLUTELY MUST', 'IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE') to override default agent behavior and force the invocation of external skill tools before any response or action.
[INDIRECT_PROMPT_INJECTION]: The skill establishes a protocol where external data is loaded into the agent's context and followed with high priority.
Ingestion points: Data enters the context via the output of the Skill tool (or platform equivalents like activate_skill, read_skill, and skill).
Boundary markers: Absent. While the instructions allow the agent to ignore a skill if it is 'wrong for the situation', there are no explicit delimiters or instructions to ignore embedded commands within the loaded content.
Capability inventory: Across the referenced files, the agent is granted access to file system operations (Read, Write, Edit), shell command execution (Bash, terminal), and network fetching (WebFetch, fetch).
Sanitization: No sanitization or validation of the content returned by external skills is performed before the agent is instructed to follow it exactly.
[DYNAMIC_EXECUTION]: The references/codex-tools.md file provides a framework for the dynamic assembly of prompts by reading template files, filling placeholders with context, and spawning new subagents using the resulting instructions.

using-superpowers