The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from user-provided LaTeX manuscripts (e.g., main.tex) and cover letter drafts (e.g., cover_letter.md). This creates a surface where malicious instructions could be embedded within the academic text to influence the agent's behavior.
Ingestion points: The agent reads main.tex and cover_letter.md using the provided analysis scripts, including scripts/tex_loader.py and scripts/extract_manuscript_facts.py.
Boundary markers: SKILL.md includes explicit safety boundaries instructing the agent to treat all extracted text as untrusted evidence rather than instructions and to ignore any embedded requests to reveal prompts or perform unauthorized actions.
Capability inventory: The skill utilizes Read, Glob, Grep, and Bash(uv *) tools. It is capable of executing its own local Python scripts but lacks network access or file-write capabilities (it explicitly avoids overwriting user files).
Sanitization: The Python scripts utilize regex-based extraction, whitespace normalization, and LaTeX-specific parsing to process data, reducing the risk of direct instruction execution.
[COMMAND_EXECUTION]: The skill leverages the uv package manager via Bash to execute local Python scripts (e.g., uv run python -B $SKILL_DIR/scripts/cover_letter.py). These executions are deterministic, pointing only to internal skill scripts, and do not involve remote code execution or the processing of unsanitized external shell commands.

cover-letter