The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell commands to execute bundled Python scripts (metrics.py, metrics_v2.py, humanize_html.py) for text analysis and processing. These scripts rely on standard libraries and perform all operations on local files within the agent's defined workspace.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided Korean text, which inherently presents a surface for indirect prompt injection if the source text contains instructions intended for the agent.
Ingestion points: The agent saves raw user input to _workspace/{run_id}/01_input.txt in Phase 1.
Boundary markers: The skill incorporates a mandatory "Do-NOT list" in Phase 3-2 and a 6-point "Self-verification" check in Phase 3-5 to prevent the agent from modifying facts, citations, or following embedded instructions that deviate from the primary task.
Capability inventory: The agent uses bundled Python scripts and the Edit tool to perform targeted text replacements.
Sanitization: While the skill does not use explicit prompt delimiters, its design relies on quantitative metric comparison (Before/After) and register preservation rules to detect and revert unauthorized changes.
[SAFE]: The skill's metadata and reference files are consistent with its stated functionality. The bundled scripts are transparent, well-documented, and do not contain obfuscated code or hidden network calls.

korean-humanize