What problem does it solve? Visitor thumbs-down verdicts on a site's AI chat accumulate in Redis but are never read, so real answer failures never become regression tests or fixes. This Skill closes that feedback loop on a biweekly cadence. ## Core Features & Use Cases - Feedback triage: Reads thumbs verdicts from the live Upstash Redis (read-only), joins them to the logged exchanges, and identifies genuine answer failures without manufacturing cases from fine answers. - Golden-set growth: Encodes each real failure as a paraphrased golden-set case with required facts and assertions per the eval spec, protecting visitor privacy by keeping verbatim questions out of commits. - Eval-driven fixes: Runs the answer-quality eval once, fixes corpus sources, persona, or chat lookup tools, and lands everything on a local branch with a plain-English report for approval — never pushing or deploying. - Use Case: Run the loop every two weeks to turn the last window of disliked chat answers into new regression cases, verify they fail before and pass after the fix, and hand the user a branch to merge or reject. ## Quick Start Ask the AI to run the chat quality loop for the feedback received since the last report.
