aeon

Trigger phrases

Phrases that activate this skill when typed to Claude Code:

• time series classification
• temporal anomaly detection
• time series clustering
• univariate time series ML

What it does

aeon is a Claude Code skill from K-Dense AI’s scientific-agent-skills repo. It turns Claude into a time-series ML specialist, covering the full range of temporal learning tasks: classification, regression, clustering, forecasting, anomaly detection, segmentation, and similarity search.

The output of a session is working Python code using aeon’s scikit-learn-compatible API — fit/predict pipelines, panel datasets, and algorithm-specific parameter choices drawn from the library’s extensive collection of time-series-native methods (ROCKET, HIVE-COTE, BOSS, etc.).

Aeon fills the gap that standard scikit-learn leaves open: it treats time as a first-class dimension rather than flattening sequential data into feature vectors.

When to use it

Reach for it when:

• You have labeled time series and want to train a classifier or regressor that respects temporal structure
• You need anomaly detection or change-point segmentation on sensor streams, physiological signals, or log data
• You want to compare multiple time-series-specific algorithms (ROCKET, WEASEL, DTW-based) under a unified API

When not to reach for it:

• Your data is tabular with no meaningful temporal ordering — standard scikit-learn is simpler
• You need deep sequence models (LSTMs, Transformers) — reach for pytorch-lightning or transformers instead

Install

Copy the SKILL.md from scientific-skills/aeon into .claude/skills/aeon/.

The skill activates on trigger phrases including “time series classification”, “temporal anomaly detection”, and “time series clustering”.

What a session looks like

A typical session has three phases:

1. Data loading. Claude shapes your raw time series into aeon’s panel format — a 3D array (samples × channels × time points) or a nested DataFrame — and handles unequal-length series if present.
2. Algorithm selection. Claude picks an appropriate estimator for the task (e.g., RocketClassifier for accuracy, MiniRocket for speed) and constructs the fit/predict pipeline with sensible defaults.
3. Evaluation. Claude generates cross-validation code using aeon’s temporal splitters, reports accuracy or F1 per class, and flags any dataset characteristics (class imbalance, variable length) that might affect results.

Receipts

Honest reporting on what aeon handles well and where it falls short:

Where it works well:

• Benchmarking multiple classifiers on UCR/UEA archive datasets with minimal boilerplate
• Anomaly detection on multivariate physiological signals where per-channel methods underperform
• Rapid prototyping of time-series pipelines before committing to a custom deep-learning stack

Where it backfires:

• Very long time series (hundreds of thousands of time points per sample) can make some algorithms impractically slow; the skill may not proactively flag this
• Forecasting support in aeon is newer and less battle-tested than the classification/clustering modules

Pattern that works: start with RocketClassifier as a baseline — it is almost always competitive and fits quickly. Then branch to task-specific algorithms only if ROCKET underperforms on your validation set.

Source and attribution

Originally authored by K-Dense, Inc.. The canonical SKILL.md lives in the aeon folder of the scientific-agent-skills repository.

License: MIT. Install, adapt, and redistribute with attribution preserved.

This page documents the skill from a practitioner’s perspective. For the formal spec and updates, defer to the source repo.