The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (CSV) and user-supplied text (abstracts, titles) to generate plotting scripts and AI-assisted schematic prompts. This creates a surface for indirect prompt injection where malicious data could influence the agent's code generation.
Ingestion points: read_csv_rows in scripts/plot_templates.py and text arguments in scripts/generate_openrouter_schematic.py.
Boundary markers: The skill implements logical boundaries through a "Figure Contract" workflow, although it lacks explicit physical delimiters for user-supplied strings interpolated into generated code.
Capability inventory: The skill possesses file-writing capabilities (savefig, write_text, makedirs) and network access for image generation.
Sanitization: Standard CSV parsing and Python Abstract Syntax Tree (AST) validation are used to preflight generated code.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to the OpenRouter image API (openrouter.ai) and supports other user-defined image generation endpoints. These are well-known services used for the skill's stated purpose of drafting manuscript schematics.
[COMMAND_EXECUTION]: The R-based audit workflow in scripts/panel_alignment.R utilizes system2 to execute the internal Python utility audit_panel_alignment.py. This is an expected inter-process communication for cross-backend figure validation.

nature-figure