The Agent Skills Directory

[SAFE]: The skill is a collection of Markdown and JSON files defining rules for agent behavior. It does not contain any executable scripts, binary files, or network-enabled commands.
[NO_CODE]: No source code is included in the skill. The logic is expressed entirely as natural language instructions and structured data.
[DATA_EXFILTRATION]: There are no patterns for accessing credentials, environment variables, or private user files, nor are there any mechanisms to transmit data to external servers.
[PROMPT_INJECTION]: The instructions establish a hierarchy for skill consumption to ensure consistency. These are not malicious injections and do not attempt to override model safety guardrails.
[EXTERNAL_DOWNLOADS]: The skill does not reference or download any external packages, dependencies, or remote scripts.
[INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to ingest and process other skills. While this constitutes an attack surface, the protocol is specifically designed to mitigate risk.
Ingestion points: The agent context where external skills are evaluated during the 'Gate' and 'Brain' stages (defined in SKILL.md).
Boundary markers: The protocol enforces explicit trust verification and adversarial evaluation by the 'Judge' to identify and block malicious instructions.
Capability inventory: No subprocess calls, network operations, or file-write capabilities are present in this skill.
Sanitization: External skills are filtered through a registry-based trust check and collision resolution logic before being admitted to the context.

The Standard Consumption Skills