skill-tasting

Output quality A/B for an existing skill — generate variants with different output traits, run them blind against the same test prompts, present side-by-side to the user, capture preference (A / B / both / neither). Each iteration is human-judged; no LLM ratchet on taste-sensitive outputs. Constitution honored as floor (variants violating MUST clauses auto-rejected before user sees them); preference log accumulates as an RLHF-lite dataset; an H4 self-trained judge replaces the LLM-judge once the log is dense enough. Use when the user wants to improve a skill's output style, quality, voice, or tone, explore variant phrasings, or accumulate preference signal for taste-sensitive skill work. Triggers: improve skill output / A/B variants / output quality / style optimization / run A/B tests / try different phrasing / taste-sensitive skill. Do NOT use for token- or structure-refactors that must preserve output (use dev-workflow:skill-refactor), structural redesigns (use dev-workflow:skill-creator-advance), creating a new skill (use dev-workflow:skill-creator-advance), or deterministic/mechanical skills where output is binary correct/incorrect (no taste dimension for A/B). Skill output A/B · human-judged · preference log.