What problem does it solve?
Running an A/B experiment tells you which variant won, but not what users actually did in each variant. This Skill provisions a Replay Vision scanner scoped to one experiment's exposed sessions, so an LLM watches session recordings and tags behavior per variant without hand-built exposure filters.
Core Features & Use Cases
- Server-side exposure scoping: Sets
experiment_targeting so the PostHog API derives the person-scoped exposure filter, covering exposures fired server-side or in earlier sessions.
- Comparable prompt templates: Provides classifier templates ("Did anyone notice?", post-exposure friction, funnel drop-off) with escape tags and variant-blind prompts so tag shares stay comparable across variants.
- Experiment-aware sizing and safe rollout: Sizes credit spend against the experiment's own population and remaining run time, creates the scanner disabled, and previews the prompt on real sessions before enabling.
- Per-variant readout: Joins
$recording_observed events to exposure events in HogQL to tally tags per variant, with guards against multi-variant sessions and broken feature-flag gates.
- Use Case: Your test variant is losing and the metrics can't say why. Use this Skill to stand up a friction classifier over exposed sessions, preview it on a few recordings, then read a per-variant tally of confusion and dead-end tags.
Quick Start
Set up a Replay Vision scanner for my running checkout experiment so I can see what users actually do in each variant.
