The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill utilizes the pickle module for local caching and checkpointing of protein embeddings and generation results. The pickle.load() function is implemented in references/esm-c-api.md (within the EmbeddingCache class) and references/forge-api.md (within the CheckpointedBatchProcessor and ForgeCache classes) to deserialize data from local files such as embeddings_cache.pkl and checkpoint.pkl. This practice involves unsafe deserialization of local data which could potentially lead to code execution if the cache files are compromised.
[INDIRECT_PROMPT_INJECTION]: The skill ingests protein sequences and structure files, which represents an attack surface for indirect prompt injection.
Ingestion points: Data enters the system via ESMProtein initialization with raw sequences and loading PDB files using ESMProtein.from_pdb(), as shown in references/esm3-api.md and references/workflows.md.
Boundary markers: The instructions do not define delimiters or include specific instructions to the model to delineate or ignore potential injection patterns within the biological data inputs.
Capability inventory: The skill's capabilities include file system write operations (PDB, FASTA, PNG), network operations via the Forge API, and execution of shell commands for library installation.
Sanitization: There is no logic provided to sanitize or validate the content of protein sequences or the structure of PDB files before they are processed by the models.
[EXTERNAL_DOWNLOADS]: The skill connects to EvolutionaryScale's Forge API at https://forge.evolutionaryscale.ai for inference tasks and retrieves model weights from HuggingFace. These external references are appropriate for the skill's stated purpose of working with ESM models.
[COMMAND_EXECUTION]: The SKILL.md document includes commands to install the esm and flash-attn Python libraries using uv pip install, which are standard installation procedures for the toolkit.

esm