The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill installs the matchms library (version 0.33.1) using the uv package manager. This is a legitimate scientific package available through standard registries.
[EXTERNAL_DOWNLOADS]: The skill interacts with recognized life-science services, including the Metabolomics-USI resolver (metabolomics-usi.gnps2.org) and PubChem, for data retrieval and metadata enrichment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and parse spectral data files (MGF, MSP, mzML, JSON) from potentially untrusted sources. These files contain metadata fields where adversarial instructions could be embedded to influence the agent's interpretation of findings.
Ingestion points: Files are loaded via the load_spectra utility and remote USI requests in SKILL.md and scripts/library_search.py.
Boundary markers: The instructions mandate a rigorous validation workflow, including QC metrics and mirror-plot inspections, which encourages human-in-the-loop oversight.
Capability inventory: The skill possesses capabilities for file system operations (read/write/edit) and network requests to scientific APIs.
Sanitization: The provided Python script filters ingested data into structured CSV outputs, which helps limit the exposure of the LLM to raw, potentially malicious metadata strings.

matchms