The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads pretrained machine learning models from Hugging Face repositories (e.g., SpikeInterface/UnitRefine_noise_neural_classifier) and refers to scientific packages hosted on GitHub and official registries. These downloads target well-known services and reputable scientific organizations.
[DYNAMIC_EXECUTION]: The skill utilizes the skops library to load serialized machine learning models for unit classification. This dynamic loading process requires the trust_model=True flag. The documentation in references/AUTOMATED_CURATION.md includes explicit security warnings, advising users to treat these models as executable artifacts and only load them from trusted sources.
[INDIRECT_PROMPT_INJECTION]: The skill supports AI-assisted curation by sending rendered plots of neural data to vision-language models. This introduces a surface where processed recording data can influence agent behavior.
Ingestion points: Data is ingested via si.read_spikeglx, si.read_openephys, and si.read_nwb (documented in SKILL.md and scripts/neuropixels_pipeline.py).
Boundary markers: The prompt used for AI curation (found in references/AI_CURATION.md) does not use explicit boundary markers to separate the visual content from the classification instructions.
Capability inventory: The skill has capabilities to execute subprocesses (via si.run_sorter), manage files, and perform network requests to AI APIs and model registries.
Sanitization: There is no evidence of specific sanitization or filtering applied to the generated visualizations before they are processed by the AI models.

neuropixels-analysis