The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in banned-phrases-ko.md and stage-flow-short.md suggest that the agent use shell commands like grep and wc to analyze user-provided text. The provided command template (echo "$paragraph" | grep ...) is vulnerable to shell command injection if the $paragraph variable contains malicious shell metacharacters.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-submitted writing samples and, in certain environments like Claude Code, may use tools to store or analyze this data. This creates a surface for indirect prompt injection where adversarial instructions hidden within the user's text could influence the agent's behavior or tool usage. * Ingestion points: User input for writing drafts during Stages 1, 2, 3, and 4 (identified in stage-flow-short.md and stage-flow-long.md). * Boundary markers: The skill uses markdown templates and specific stage-based triggers, but lacks explicit instructions to ignore instructions embedded in the user text. * Capability inventory: Potential use of write_file tool to save progress (e.g., proposal.md) and potential use of bash for text analysis as described in the README and reference files. * Sanitization: No explicit sanitization or escaping of user-provided content before processing or interpolation into commands is mentioned.

academic-writing-trainer