Skill UI

Activation-aware weight quantization that compresses 7B-70B models to 4-bit with minimal accuracy loss, delivering ~3x inference speedups and memory reductions for constrained GPU deployments while integrating with vLLM and Marlin kernels.

Enables Claude to configure neural network architectures through the neural-network-builder plugin, creating or modifying models and defining layers, activation functions, and training parameters whenever network design or tuning is requested.

Guides researchers through nnsight’s tracing and activation APIs so you can inspect or modify PyTorch internals locally or via NDIF-powered remote runs on 70B+ foundation models without local GPUs.

Pyvene is a declarative framework designed for performing causal interventions on PyTorch neural networks. It allows researchers to conduct advanced experiments such as activation patching, causal tracing (ROME-style), and interchange intervention training (IIT). Use this library when you need to test causal hypotheses about model behavior, deeply interpret model components, or ensure reproducibility in advanced AI research.

SAELens provides a framework for training and analyzing Sparse Autoencoders (SAEs). SAEs decompose the dense, often polysemantic activations of large language models into sparse, monosemantic features. Use this when you need to discover the discrete, interpretable concepts a model has learned, study feature superposition, or analyze specific safety-relevant behaviors (like bias or deception) within deep neural networks.

TransformerLens is the standard library for performing mechanistic interpretability research on large language models. It provides clean interfaces to inspect and manipulate transformer internals—such as attention patterns, residual streams, and MLP outputs—via activation caching and HookPoints. This tool is essential for reverse-engineering model algorithms, performing causal tracing, and studying the internal circuits responsible for model behavior.

This skill enables the creation and configuration of specialized Cascade agents tailored for specific domains or tasks. Users can define system prompts, integrate domain knowledge, and set precise activation triggers. It facilitates building highly customized, domain-specific AI agents and automated workflows within the Windsurf ecosystem.

Soft Predict Future is a structured foresight engine that triggers whenever the user asks about future outcomes or scenarios. It enforces a nine-step analytical pipeline (validation, signal collection, scoring, driver extraction, etc.), highlights required searches and inputs, and directs the user to web-enabled reasoning for predictive analysis.

A comprehensive coaching layer designed to elevate users into power users of AI models. Upon first activation, it provides a structured glossary of high-impact techniques tailored to user needs. In subsequent turns, it actively monitors prompts, surfacing single, actionable tips to optimize interactions and maximize the AI's capabilities, ensuring continuous, guided learning.

This skill provides comprehensive tools to detect data poisoning and model backdooring attacks throughout the machine learning lifecycle. It checks data provenance, detects label-flip anomalies (using Cleanlab), and analyzes model activations and spectra (using ART) to identify malicious triggers or compromised weights, ensuring ML supply chain integrity.