The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided academic paper source files (LaTeX or PDF) to extract text, equations, and figures. This data enters the agent's context during content planning and automated review phases. A maliciously crafted paper could attempt to inject instructions to influence the agent's behavior during these loops.
Ingestion points: Paper source ingestion in Phase 1 (SKILL.md).
Boundary markers: Present. The skill mandates fresh threads for content audits (mcp__codex__codex) and directs models to review file paths rather than inline summaries.
Capability inventory: Subprocess execution and file system writes are present across the gate scripts.
Sanitization: The skill relies on cross-model auditing to verify the fidelity of extracted claims against the source paper.
[EXTERNAL_DOWNLOADS]: The skill fetches assets from well-known external sources to support its primary functionality. It downloads the MathJax library from the jsDelivr CDN (https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-svg.js) for local rendering of mathematical formulas. It also uses web search and fetch tools to retrieve current poster specifications from official conference domains.
[COMMAND_EXECUTION]: The orchestration script (scripts/run_gates.py) and PDF verification module (scripts/_posterly/verify_final.py) utilize the subprocess module to execute internal Python scripts and system utilities like pdfinfo. The implementation uses list-based command construction with resolved absolute paths to mitigate potential shell injection vulnerabilities.

paper-poster-html