The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill identifies and references official model weights and configurations from the Biohub organization's Hugging Face repositories (e.g., biohub/ESMC-300M, biohub/ESMFold2). These are standard scientific resources and are documented neutrally as required for well-known research organizations.
[COMMAND_EXECUTION]: Provides standard environment setup instructions using the uv package manager to install the verified esm library from the official Python Package Index (PyPI). It uses pinned versions to ensure supply chain consistency.
[INDIRECT_PROMPT_INJECTION]: The skill describes an attack surface related to processing external biological sequence data and protein structure files.
Ingestion points: Untrusted inputs include protein sequences provided in string format and structural data loaded from PDB/mmCIF files via ESMProtein.from_pdb. It also loads model weights from external Hugging Face repositories.
Boundary markers: The instructions explicitly advise rejecting accidental gaps, spaces, or FASTA headers from input sequences to prevent malformed data ingestion.
Capability inventory: The skill possesses capabilities to write files (to_pdb, to_mmcif) and perform network requests for hosted inference via the esm SDK clients directed at https://biohub.ai.
Sanitization: The skill includes a dedicated helper script scripts/esm_embeddings.py which contains a validate_sequence function. This function strictly validates input characters against the known amino acid alphabet and enforces length constraints before processing data.
[SAFE]: The skill follows secure credential management by instructing the use of os.environ["ESM_API_KEY"] rather than hardcoding secrets. All network operations are directed to the service's official domain (biohub.ai).

esm