The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, primarily through 'revision_coach_agent.md' (reviewer comments) and 'literature_strategist_agent.md' (literature search abstracts). Mandatory evidence chain: 1. Ingestion points: 'revision_coach_agent.md' parses free-form reviewer feedback; 'literature_strategist_agent.md' screens external database results. 2. Boundary markers: The pipeline uses structured XML-like tags (e.g., phase-output delimiters) and invisible HTML comments (e.g., ref and anchor markers) to separate user data from instructions. 3. Capability inventory: Access to shell commands for document formatting (Pandoc/LaTeX) and Python/R environments for visualization. 4. Sanitization: Extensive 'IRON RULE' markers, structural lints, and 'Writing Quality Check' diagnostics enforce integrity and prevent instruction overrides.
[DYNAMIC_EXECUTION]: The 'visualization_agent.md' generates Python (matplotlib/seaborn) and R (ggplot2) code based on quantitative findings in the Results section. The 'vlm_figure_verification' protocol describes an automated step to execute this generated code to render images for quality auditing. This capability is standard for the skill's primary research purpose.
[COMMAND_EXECUTION]: The 'formatter_agent.md' and the 'revision_patch_protocol' recommend the use of standard shell commands for Pandoc, LaTeX (tectonic, xelatex), and various local Python utility scripts. These commands are appropriately scoped to document conversion and pipeline integrity tasks.

academic-paper