The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, such as page HTML and user-provided briefs, which constitutes an attack surface for indirect instructions.
Ingestion points: Defined in SKILL.md as page HTML, source material, and the user's brief.
Boundary markers: The skill includes a dedicated 'Security boundary' section that explicitly labels input as untrusted data and mandates that it never be treated as instructions. It specifically prohibits input from authorizing tools or overriding core rules like the 'no-invention' constraint.
Capability inventory: The skill is restricted to text generation and does not have access to subprocesses, shell execution, or network tools.
Sanitization: Implements a strict 'no-invention rule' and a mandatory output contract that requires all generated facts to be traced back to the user, effectively filtering out hallucinated or injected claims.
[NO_CODE]: The skill contains no executable scripts, binaries, or automated installers, consisting only of markdown instructions and metadata.

awesome-copywriting