Salesforce Engineering shares how they built a self-improving AI system using automated feedback loops. The core idea: repeated code review comments are undocumented requirements, so they built a meta-skill generator that encodes reviewer wisdom into the system itself. Quality is measured via a three-tier evaluation framework — trigger accuracy (tested against 80+ competing skills), structural validation (21 deterministic rules), and rubric-based LLM-as-judge scoring. A weekly automated workflow mines PR comments, deduplicates patterns by frequency × severity, applies surgical edits (capped at 5 improvements and 100 lines per cycle), and validates changes before opening a draft PR. Over six cycles, the system converged naturally — applying changes in only two cycles and self-limiting in four. When frontmatter conventions changed between cycles 6 and 7, a frequency spike in reviewer comments triggered automatic re-adaptation without manual intervention. The same architecture was applied to an automated code reviewer bot, teaching it what humans actually care about. Prerequisites for this pattern: sufficient review volume (15+ merged PRs per cycle), structured verifiable output, and a meaningful review culture.
