Preference Skills
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Preference Skills are a category of model fine-tuning or skill-teaching that encodes relatively stable behavioral preferences rather than raw capabilities, distinguished by their durability across model generations.
Definition and Characteristics
Philipp Schmid (Google DeepMind) introduces Preference Skills as one of two fundamental skill categories when building on top of language models. In contrast to Capability Skills — which teach models something they "cannot do consistently at the moment" and are expected to become obsolete as base models improve — Preference Skills "are more durable, mostly encode some references." →
Role in Skill Lifecycle and Evals Strategy
The durability distinction between the two skill types carries direct implications for how teams should invest in evals. Because Capability Skills are temporary by nature (they fill gaps that frontier model improvements will eventually close), they warrant lighter-touch evaluation and should be retired when no longer needed. Preference Skills, being durable, are the ones that should be actively protected with robust evals: they represent intentional product or alignment choices that should not erode across model updates or fine-tuning iterations. →
Context
This framework is presented in Schmid's talk Don't Ship Skills Without Evals, where the core argument is that shipping any skill — but especially durable Preference Skills — without corresponding evaluation coverage is a reliability risk. The two-category model gives practitioners a principled way to prioritize eval investment: protect what is meant to last.