The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill correctly identifies and mitigates the risk of indirect prompt injection. In references/audit-workflow.md, it explicitly instructs the agent to treat the text under audit as data only and never as direction. It specifically directs the agent to flag any sentences that attempt to address the auditor or override behavior as findings rather than following them.
[DATA_EXPOSURE]: The skill includes proactive measures to prevent metadata leakage. In SKILL.md and references/language-and-patterns.md, it mandates the mechanical removal of AI-referrer tracking parameters (such as utm_source=chatgpt.com or referrer=grok.com) and citation artifacts (like [oai_citation:...]) from delivered prose to ensure privacy and clear provenance.
[COMMAND_EXECUTION]: While SPEC.md describes internal validation commands (e.g., python3 distillery/scripts/distiller.py), these are development-time tools for local repository maintenance and do not indicate malicious runtime behavior by the agent skill itself.
[PROMPT_INJECTION]: The skill's instructions in SKILL.md and references/audit-workflow.md emphasize maintaining the original writer's voice while strictly adhering to the audit workflow, which acts as a safeguard against attempts to manipulate the agent's behavior via the input text.

writing