The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions are purely stylistic and operational, defining a specific creative 'vector' for the agent. There are no attempts to override safety filters, bypass instructions, or extract system prompts. The framing used is intended to guide output quality rather than subvert the underlying model behavior.
[DATA_EXFILTRATION]: No network operations (such as curl or fetch) or access to sensitive system paths (e.g., .ssh, .aws) are present. The skill focuses on reading and editing local project files identified by the user.
[OBFUSCATION]: A thorough scan for Base64 encoding, zero-width characters, homoglyphs, and hidden text patterns revealed no obfuscated content or hidden URLs.
[COMMAND_EXECUTION]: While the skill's configuration allows access to the Bash tool, the instructions do not contain or generate dangerous shell commands. It primarily uses file-reading commands (Grep, Glob) for gathering context from the project, which is consistent with its stated purpose of editing text.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted text from external files for processing. However, it incorporates a 'neuroslop' pass through an internal writer tool as a final safety and quality check.
Ingestion points: Reads project files (e.g., via audit <file> or rewrite <file:start-end>).
Boundary markers: Not explicitly defined in the instructions for text interpolation.
Capability inventory: Includes Write, Edit, Bash, Grep, and Glob tools.
Sanitization: Employs a final pass through an external writer skill to check for undesirable patterns.
[DYNAMIC_CONTEXT_INJECTION]: No !command syntax or load-time shell execution patterns were found in the skill's markdown frontmatter or body.

pelevin-digression