The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external, potentially untrusted data such as SMILES strings and protein structures which are then used as inputs for complex molecular modeling workflows that involve both network operations and file system writes.
Ingestion points: SMILES strings, protein sequence strings, and PDB/mmCIF files are ingested through functions like rowan.Molecule.from_smiles(), rowan.upload_protein(), and submit_protein_cofolding_workflow() in SKILL.md and workflow_catalog.md.
Boundary markers: The instructions note that the rowan-python SDK performs validation on SMILES inputs and protein file syntax, providing some level of structural boundary enforcement.
Capability inventory: The skill has the capability to perform network requests (Rowan API communication), write files to the local disk (result.best_pose.write(), result.download_files()), and manage project/folder structures via the API.
Sanitization: While the SDK validates the chemistry-specific syntax of the inputs, there is no explicit sanitization for prompt injection patterns within the molecular data fields before they are processed by the cloud models.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the rowan-python package from public registries. This is the official SDK for the Rowan platform and is considered a legitimate dependency for the skill's stated purpose.

rowan