The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content (Markdown, HTML, and text files) to perform reviews. This presents an indirect prompt injection surface where instructions hidden in the text could attempt to influence the agent's persona or verdict.
Ingestion points: Untrusted data enters the agent context via git diff outputs and file reads performed in SKILL.md.
Boundary markers: The instructions do not specify explicit delimiters or "ignore embedded instructions" warnings for the content being reviewed.
Capability inventory: The skill utilizes Read, Grep, Glob, Bash (git), and executes a local Python script (scripts/skan_pisowni.py).
Sanitization: The provided Python script uses HTMLParser to strip tags and applies regular expressions for pattern matching, though it does not provide specific prompt-level sanitization.
[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/skan_pisowni.py) included in its package. Analysis of the script confirms it only uses standard libraries (json, re, sys, html.parser, pathlib) and performs text processing without dynamic code execution or external network calls.
[SAFE]: The skill is designed for local content review, lacks network capabilities for data exfiltration, and limits its tool usage to project-related file operations. Its instructions strictly define a limited output format and persona, reducing the impact of potential prompt manipulations.

marko-pl-content