← ClaudeAtlas

taste-enginelisted

Turn subjective quality intent into executable hard gates, weighted rubrics, exemplars, anti-exemplars, blind comparison, and independent judge evidence. Use for high-value artifacts where correctness alone is insufficient.
frankxai/Starlight-Intelligence-System · ★ 9 · AI & Automation · score 72
Install: claude install-skill frankxai/Starlight-Intelligence-System
# Taste Engine ## Outcome Create a domain-specific Taste Profile and an evaluation loop that can reject polished mediocrity without pretending subjective judgment is deterministic. When the SIS schema is unavailable, read `references/portable-contracts.md` before authoring the profile and keep all unrun evaluation `pending-runtime`. ## Use this skill when - A high-value artifact must be persuasive, elegant, clear, resonant, or production-ready. - “Premium,” “beautiful,” “world-class,” or similar intent needs an executable definition. - Multiple candidates can be compared. - A winning example should become reusable preference memory. Do not use a taste rubric as a substitute for factual, security, accessibility, or artifact-native checks. ## Procedure 1. Identify the judgment context. - Name the artifact type, audience, decision, medium, production constraints, and what failure costs. 2. Write hard rejection gates. - Encode failures that invalidate an artifact regardless of polish. - Keep factual fabrication, unusable output, accessibility failure, and broken production constraints outside weighted averaging. 3. Define weighted dimensions. - Use observable descriptions of excellence and failure. - Make weights reflect the use case rather than generic aesthetic preference. - Confirm weights total approximately 1.0. 4. Ground the rubric. - Add reference exemplars and explain why each wins. - Add anti-exemplars and name the failure pattern. - Do