The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external text provided by users (in edit mode) or briefs (in write mode) to perform content optimization. This input is interpolated into the agent's context without explicit boundary markers, creating a potential surface for indirect prompt injection if the processed content contains malicious instructions.
Ingestion points: Untrusted data enters the agent context through the TEXT, BRIEF, and SAMPLE input variables defined in SKILL.md.
Boundary markers: The skill procedure lacks instructions to use delimiters or 'ignore' instructions to isolate the user-provided text from the system prompt.
Capability inventory: The skill possesses the capability to write the processed output to files in the working directory and perform recursive 'audit' calls, which could be influenced by injected instructions.
Sanitization: There is no mention of sanitizing, escaping, or validating the input text before it is processed by the LLM.
[DATA_EXPOSURE_&_EXFILTRATION]: No sensitive data access or network exfiltration patterns were detected. All URL references are to reputable sources such as Wikipedia and GitHub for documentation and methodology purposes.
[REMOTE_CODE_EXECUTION]: The skill does not perform any remote code downloads or executions. It relies on internal logic and local reference files.
[COMMAND_EXECUTION]: The skill does not execute shell commands or spawn subprocesses, beyond a standard file-write operation for its output.

concise-writing