The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's UnitRefine curation functionality utilizes the 'skops' library with the 'trust_model=True' parameter to load pretrained classifiers. This code pattern enables the deserialization of arbitrary Python objects from models fetched from Hugging Face, bypassing standard security filters designed to prevent the loading of potentially malicious serialized data.
[COMMAND_EXECUTION]: The skill uses the 'SpikeInterface' API ('si.run_sorter') to execute external spike sorting software such as Kilosort4 and SpykingCircus2. These tools are invoked as subprocesses and can be configured to run within Docker or Singularity containers, granting the agent the ability to execute external binaries and manage container environments.
[EXTERNAL_DOWNLOADS]: The skill fetches machine learning models from the official 'SpikeInterface' repository on Hugging Face. Additionally, the documentation provides instructions for installing and referencing scientific software hosted on third-party GitHub repositories, including repositories for Kilosort, Bombcell, UnitMatch, and DREDge.
[DATA_EXFILTRATION]: An optional AI-assisted curation workflow transmits unit summary images to external vision-language model APIs (Anthropic and OpenAI). The skill correctly implements security best practices by using environment variables ('os.environ') for API key management, ensuring credentials are not hardcoded in analysis scripts.
[INDIRECT_PROMPT_INJECTION]: The skill processes large external datasets and remote models, creating a vulnerability surface. Ingestion points: Raw recording files (SpikeGLX, Open Ephys, NWB) and Hugging Face model weights. Boundary markers: Absent. Capability inventory: Subprocess execution via external sorters and network API calls for curation. Sanitization: Absent (validation steps focus on data integrity and format compliance rather than malicious content).

neuropixels-analysis