steering

name

steering

description

Run activation-steering feature discovery for a Hugging Face model. Use this skill whenever the user says /steering, wants to steer a model, extract a cognitive feature, generate steering vectors, run feature discovery, or produce steering artifacts. Specify a model (default: gpt2) and an optional feature name. If a feature is supplied, generate inputs, expected outputs, then run an extraction pass and output artifacts. If no feature is supplied, auto-pick one that hasn't already been extracted. If no test data is supplied, generate synthetic examples; otherwise use the user's data and fill in whatever's missing.

/steering

Run activation-steering feature discovery end-to-end: specify a model and an optional feature, and this skill orchestrates data generation, vector extraction, and artifact output.

When to use this skill

• The user types /steering or asks to "steer" a model
• The user wants to extract a cognitive feature (chain_of_thought, few_shot_prompting, react, etc.)
• The user wants to generate steering artifacts for a model
• The user asks to run feature discovery or build a steering vector
• The user mentions activation engineering, representation engineering, or contrastive extraction

How it works

Inputs

Parameter	Default	Description
model_name	"gpt2"	Any decoder-only Hugging Face model identifier
feature_name	auto-pick	A feature to extract (e.g. chain_of_thought). If omitted, the system picks the next undiscovered feature from the standard catalog.
user_examples