The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its heavy reliance on external scientific data sources.
Ingestion points: The workflow (SKILL.md) explicitly instructs the agent to use 'WebFetch' for PubMed URLs and 'WebSearch' for broader scientific domains in Step 2.
Boundary markers: There are no instructions to use XML tags, specific delimiters, or 'ignore embedded instructions' warnings for the data retrieved from external sources.
Capability inventory: The agent has access to 'Bash' (for script execution and LaTeX compilation), 'Write', and 'Edit' tools.
Sanitization: The instructions do not specify any validation or sanitization of external text before it is synthesized into hypotheses or interpolated into LaTeX templates.
[COMMAND_EXECUTION]: The skill relies on local shell execution for core functionality.
Evidence: The skill instructs the agent to execute python scripts/generate_schematic.py with natural language descriptions as arguments. It also requires the execution of xelatex and bibtex to produce PDFs.
Context: While these commands are standard for scientific document preparation, they represent a surface for command injection if malicious strings are ingested from external literature and passed directly to these CLI tools.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to retrieve scientific literature and may involve package installation.
Evidence: Step 2 of the workflow uses 'WebFetch' on PubMed and general web searching. Additionally, assets/FORMATTING_GUIDE.md suggests using tlmgr install for missing LaTeX packages.
Context: These operations target well-known scientific databases and official TeX registries, which are generally low-risk but involve the ingestion of third-party content.

hypothesis-generation