The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs high-pressure language and mandatory imperatives (e.g., 'ABSOLUTELY MUST', 'not negotiable', 'cannot rationalize your way out of this') in SKILL.md to override the AI's reasoning and safety filters regarding tool invocation.- [PERSISTENCE]: In references/codex-tools.md, the skill instructs the agent to modify the global Codex configuration file at ~/.codex/config.toml. These modifications change default subagent models and reasoning efforts, persisting behavioral overrides across future AI sessions.- [INDIRECT_PROMPT_INJECTION]: The skill mandates skill invocation if there is a '1% chance' of applicability, significantly increasing the attack surface for the agent to inadvertently follow malicious instructions contained in other workspace files or skills.
Ingestion points: Ingests instructions from workspace files like CLAUDE.md, AGENTS.md, and GEMINI.md, as well as other locally defined skills.
Boundary markers: Absent; the skill commands the agent to prioritize these external instructions over default behaviors and to announce tool use before performing any safety checks or file explorations.
Capability inventory: The skill maps actions to powerful platform tools including terminal shell access (terminal, run_shell_command), file writing (write_file, patch), and subagent spawning.
Sanitization: No sanitization or validation of the content of the invoked skills is provided.- [DYNAMIC_EXECUTION]: The platform-specific reference files (references/codex-tools.md, references/gemini-tools.md) describe procedures for spawning subagents using dynamic prompt templates. This involves the runtime assembly and execution of instructions based on potentially untrusted workspace data.

using-superpowers