The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/prose_wordcount.py executes the system utility detex to process LaTeX content.
Evidence: subprocess.run(["detex", "-n"], input=text, capture_output=True, text=True) in scripts/prose_wordcount.py. The implementation uses a fixed argument list and avoids shell invocation, minimizing injection risk from command arguments.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and interpret user-provided manuscripts, which creates an attack surface for indirect prompt injection.
Ingestion points: scripts/prose_wordcount.py reads local .tex files and recursively resolves \input and \include directives to build the full text of the manuscript.
Boundary markers: The skill does not provide the agent with explicit instructions or delimiters to distinguish between manuscript text and administrative instructions during audit passes.
Capability inventory: The skill provides logic for the agent to execute shell commands via the word count script and perform automated text substitutions (replace_exactly) across files.
Sanitization: The resolve_inputs function does not validate or restrict the file paths provided in \input directives, which could be exploited to read sensitive files if a manuscript contains malicious path references.

draft-marker-discipline