The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text provided by users to perform 'de-slopping' (removing AI-writing fingerprints). This creates a vulnerability where malicious instructions embedded in the input text could be interpreted by the agent during the rewrite phase.
Ingestion points: User-provided text strings or temporary files processed in Phase 1 (Mechanical scan) and Phase 2 (Rewrite).
Boundary markers: The instructions do not define clear delimiters (e.g., XML tags or unique markers) to separate user data from agent instructions.
Capability inventory: The skill uses grep for pattern matching and performs iterative text rewriting, which could be exploited if the agent follows instructions found within the 'slop' text.
Sanitization: No validation or sanitization of the input text is specified before it is processed or rewritten.
[COMMAND_EXECUTION]: Phase 1 instructions in SKILL.md and references/tells.md direct the agent to execute shell commands (specifically grep, tr, sed, and awk) against local files containing user-supplied text. While these are standard utilities, this establishes a pattern of executing commands directly on untrusted file content.

z-fuck-slop