Where Rule-Based (Objective) Agent Output Verification handles objective pass/fail checks, taste-based verification judges qualities that have no binary test — design quality, the quality of a content idea, writing concision — by configuring the AI itself to apply a defined "taste" or quality bar.
Unlike rule-based checks, a taste-based verifier is never really finished: it's built as a Claude Skill that encodes the operator's subjective standard, then iterated on over time as edge cases surface.
The presenter's own worked example is an "anti-slop" skill (built on the buildpartner.ai Claude plugin, linked in the source video's description) — a taste-based verifier whose specific job is catching generic, low-effort AI output ("slop") before it ships.
Apply: Build a second Claude Skill, separate from your rule-based one, that encodes your subjective quality bar. Feed it examples/context and revise it over time as you see what it misses — treat it the way the presenter treats their anti-slop skill.