The Agent Skills Directory

[SAFE]: The skill consists of markdown documentation and examples designed to improve prompt engineering. It does not contain any executable code, scripts, or network operations.
[COMMAND_EXECUTION]: The documentation references the use of an exec_command tool for shell operations. These references are part of the guidance for structuring tool contracts and include recommendations for risk-tiering, such as requiring user approval for high-risk actions.
[PROMPT_INJECTION]: The skill provides patterns and rules to defend against prompt injection. It instructs developers to validate and sanitize untrusted inputs and treat external content as untrusted data.
[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture for agents that process untrusted data.
Ingestion points: Reading external documentation, web content, or local files as mentioned in reference.md.
Boundary markers: The skill recommends using wrap_as_data and explicit delimiters to isolate data from instructions (documented in examples.md).
Capability inventory: References exec_command (shell execution) and apply_patch (file writing) in SKILL.md.
Sanitization: Recommends the implementation of validate_input and filter_output functions in examples.md and reference.md.

codex-prompt-engineering