The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses the skops library via spikeinterface.curation.model_based_label_units to load pre-trained machine learning models. The instructions specify that trust_model=True is required to load these models, which involves deserializing artifacts. This mechanism is used to fetch and apply official UnitRefine models from the SpikeInterface organization on Hugging Face.
[INDIRECT_PROMPT_INJECTION]: The skill implements AI-assisted curation by rendering visual summaries of neural units (waveforms, ISI histograms, etc.) and sending them to vision-language models for classification.
Ingestion points: Raw recording data in SpikeGLX, Open Ephys, or NWB formats processed by the SortingAnalyzer in scripts/neuropixels_pipeline.py and references/AI_CURATION.md.
Boundary markers: The skill does not implement specific boundary markers or sanitization within the visual prompts sent to AI models.
Capability inventory: The pipeline includes file system writes (analyzer.save), external binary execution via si.run_sorter, and network operations to cloud AI providers.
Sanitization: Visual data derived from raw recordings is sent directly to AI models without specific sanitization against adversarial patterns.
[EXTERNAL_DOWNLOADS]: The skill documentation and scripts facilitate the download of scientific software and models from external repositories. It fetches configuration and machine learning models from the SpikeInterface organization's official Hugging Face repository. It also provides instructions for cloning and installing research tools from public GitHub repositories, including Julie-Fabre/bombcell and EnnyvanBeest/UnitMatch.
[COMMAND_EXECUTION]: The spike sorting process (si.run_sorter) involves executing external scientific software packages (such as Kilosort4, Mountainsort5, etc.), which may run as local binaries or within Docker/Singularity containers as described in the spike sorting documentation.

neuropixels-analysis