The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its AI-assisted curation workflow. (1) Ingestion points: Raw recording data is loaded via si.read_spikeglx() and other readers in scripts/neuropixels_pipeline.py. (2) Boundary markers: The LLM prompts in references/AI_CURATION.md lack explicit instructions to ignore text or commands embedded within visual plots. (3) Capability inventory: The skill uses vision-language models (Claude/GPT) to analyze plots and return classification labels which drive further analysis steps. (4) Sanitization: No sanitization or filtering is performed on the visual summary images rendered from raw recordings before they are processed by the LLM.
[EXTERNAL_DOWNLOADS]: The documentation and scripts recommend downloading and installing third-party research tools from non-standard sources. Evidence: Instructions in references/AUTOMATED_CURATION.md include git clone commands for the bombcell and UnitMatch repositories hosted on individual GitHub accounts.
[REMOTE_CODE_EXECUTION]: The skill uses potentially unsafe deserialization to load machine learning models for unit classification. Evidence: references/AUTOMATED_CURATION.md and references/standard_workflow.md demonstrate using skops and sc.load_model with trust_model=True or trusted=['numpy.dtype'] to load .skops and .pkl artifacts from Hugging Face.

neuropixels-analysis