The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits untrusted user-supplied text, creating a potential surface for indirect prompt injection.
Ingestion points: The skill audits files, pasted text, and git diffs provided by the user, as defined in the Audit mode workflow in SKILL.md and references/audit-mode.md.
Boundary markers: The instructions include explicit safeguards, mandating that the agent "not restructure the user's own words when quoting them back" and ensuring that "Direct quotes preserved" and "Code untouched" during the audit process.
Capability inventory: The skill is composed of instructional content and does not invoke tools for network communication, subprocess execution, or arbitrary file system writes.
Sanitization: While no technical sanitization of input text is performed, the skill's operations are restricted to generating text reports and applying style edits.
[NO_CODE]: The skill consists exclusively of Markdown documentation and reference files. It does not include any Python scripts, Node.js modules, or other executable binaries.
[SAFE]: No malicious patterns, such as credential theft, persistence mechanisms, or obfuscation, were detected. The skill's behavior is consistent with its stated purpose of prose editing and auditing. The metadata and reference links correctly point to the author's own resources and established sources like Wikipedia.

anti-ai-prose