Sort spikes from a Neuropixels recording end-to-end

Hand Claude Code a raw SpikeGLX, Open Ephys, or NWB Neuropixels recording; get back motion-corrected data, Kilosort4 (or SpykingCircus2 / Mountainsort5) spike-sorted units, and an AI-assisted curation pass that flags merges, splits, and likely noise clusters before you hand-curate.

   
Problem class Data analysis
Subject areas Neuroscience
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Workstation with GPU

Problem

Neuropixels recordings yield gigabytes per probe per session and the spike-sorting pipeline that turns voltage into single units is fiddly: phase-shift correction, common-average referencing, bad-channel detection, drift estimation, sorter choice, post-sorting curation. Every lab rolls its own SpikeInterface boilerplate and the parameters that work for an Allen-style cortical recording often fail on a brainstem session. Solved looks like: a single conversation that loads the recording, applies the standard preprocessing chain, runs a state-of-the-art sorter under the right preset, and emits a curated unit table you can open in Phy or push to NWB.

  1. Install the neuropixels-analysis skill in Claude Code:

    /plugin marketplace add K-Dense-AI/claude-scientific-skills
/plugin install neuropixels-analysis@claude-scientific-skills

    Verify the environment has SpikeInterface installed (pip install spikeinterface[full]) and a CUDA-capable GPU available if you intend to run Kilosort4.

  2. Point the skill at your recording. A minimal prompt:

    Run the neuropixels-analysis skill on
data/session2026-05-01_g0_imec0/. The recording is SpikeGLX
Neuropixels 1.0, ~90 minutes, mouse motor cortex. Apply the
standard preprocessing chain: phase-shift correction, highpass
300 Hz, bad-channel detection (kurt threshold), common-average
reference. Then estimate motion with the nonrigid_accurate preset
and write the corrected recording to data/sess_preproc/.

  3. Sort with Kilosort4 (GPU) or SpykingCircus2 (CPU fallback). Continue the conversation:

    Run Kilosort4 on the preprocessed recording with the default
Neuropixels 1.0 parameters. Save sorting output to
data/sess_ks4/. Then compute SpikeInterface quality metrics
(ISI violations, presence ratio, amplitude SNR, refractory period
violation) on all units.

  4. Use the AI-assisted curation pass the skill bundles to flag obvious noise clusters and merge / split candidates:

    Use the skill's curation helpers to triage units into "good",
"mua", and "noise" categories based on the quality metrics. For
units flagged as candidate merges, show me the cross-correlograms
and waveform similarity before I confirm.

  5. Export to NWB or Phy for hand-curation. The skill writes a standard SpikeInterface sorting object that loads directly into Phy or pynwb. Hand-curation remains a human step; the skill does not claim to replace it.

  6. (Optional) Cross-sort comparison. When the result is high-stakes (a publication figure), run a second sorter (SpykingCircus2 or Mountainsort5) and ask Claude to report unit agreement via SpikeInterface.comparison. Disagreement is informative even when neither sorter is “right”.

Why this assembly

Rung 2. The skill is a thin but valuable wrapper: it encodes the SpikeInterface / Allen Institute / International Brain Laboratory consensus on preprocessing and gives Claude the right knobs for drift correction and sorter selection without re-inventing the parameter file. Plain Claude Code can write SpikeInterface code but does not reliably pick the right drift preset for the session or the right quality-metric thresholds. There is no need to escalate to an autonomous system — spike sorting is a well-defined pipeline with discrete steps and human-in-the-loop curation at the end.

Availability

Fully open. The neuropixels-analysis skill is community OSS distributed via the K-Dense marketplace; SpikeInterface, Kilosort4, SpykingCircus2, and Mountainsort5 are all open-source. No institutional license. Any current Claude plan works.

Compute requirements

Workstation with GPU. Kilosort4 effectively requires a CUDA GPU with ≥8 GB VRAM; a 90-minute Neuropixels 1.0 recording sorts in roughly 20–40 minutes on an RTX 4090. SpykingCircus2 and Mountainsort5 fall back to CPU but are 5–10× slower on the same recording. Disk: budget 3–5× the raw recording size for preprocessed and intermediate files. System RAM 32 GB recommended.

Evidence

Reported. The K-Dense neuropixels-analysis skill is the documented assembly and tracks SpikeInterface best practices directly. The closest peer-style evaluation of LLM-driven spike sorting is the SpikeAgent preprint (posted 2025-02-11), which reports a multimodal LLM-based system automating spike-sorting across recording platforms including Neuropixels, with performance varying by backend model — the recipe inherits that “model-dependent” caveat. No peer-reviewed benchmark of this exact skill against a hand-tuned SpikeInterface script is known.

The component-level evidence for the underlying tools is robust: SpikeInterface (Buccino et al., eLife 2020) and Kilosort4 (Pachitariu et al., Nature Methods 2024) are widely adopted in the systems-neuroscience community.

Alternatives considered

  • Plain Claude Code + raw SpikeInterface. Works for users who already know SpikeInterface idioms cold. The skill saves you from re-learning the parameter knobs each session and standardizes the preprocessing chain.
  • A standalone Kilosort4 GUI workflow with no Claude in the loop. The right answer when you sort the same probe layout daily and have a saved parameter file. Reach for the skill when sessions vary (different probe types, brain regions, durations) or when you want the metric-driven triage step automated.
  • Neurosift / AIND for finding recordings, then this skill for sorting. Pair the Neurosift Tools MCP or AIND Data MCP upstream when you are sorting a public DANDI dataset; this recipe assumes the recording is local.
  • An autonomous-science system. No such system targets spike sorting end-to-end today. SpikeAgent is the closest published prototype and is not a maintained installable system yet.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.