The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted research materials, such as idea.md and experimental_log.md, which are used to generate the final manuscript sections.
Ingestion points: Files in workspace/inputs/ are read by the host agent and passed to reasoning sub-tasks.
Boundary markers: The skill mandates the use of a "Universal Anti-Leakage Prompt" (from arXiv:2604.05018) to strictly isolate the agent's reasoning to the provided materials and prevent training-data leakage.
Capability inventory: The pipeline involves file writing, shell execution for LaTeX compilation, and network access for literature discovery.
Sanitization: Multiple deterministic scripts (anti_leakage_check.py, claim_evidence_gate.py) act as post-generation quality gates to detect unauthorized metadata (authors/affiliations) and verify that quantitative claims are supported by the provided experimental logs.
[COMMAND_EXECUTION]: The skill instructs the host agent to execute various Python scripts and shell commands (latexmk, pdflatex) to initialize the workspace, validate inputs, and compile the final PDF. These operations are transparent and necessary for the stated purpose of paper orchestration.
[DYNAMIC_EXECUTION]: The script scripts/check_tex_packages.py performs dynamic environment discovery by generating minimal LaTeX files and executing pdflatex to verify the availability of specific packages (e.g., cleveref, booktabs). This prevents the pipeline from failing later during the final manuscript compilation.

paper-orchestra