realistic-prompt-generation

Provide a sanitized blind design brief and prompt architecture, then generate two core variants per cell: the closest natural rendering of observed language and a natural paraphrase. Each candidate records its transformation type, evidence grade, and locale review status.

A target-blind brief contains only anonymized segments, approved jobs, constraints, journey states, locales, and evidence IDs. It excludes target brands, products, current AI answers, rankings, and the contamination register so generated prompts cannot leak target knowledge.

Only within band rules: B0 cells may name the supplied target alias in an aided pass, and B2 cells may name an evidence-supported category. B3-B5 cells supply the problem, goal, or market context without introducing products or categories not entailed by evidence.

LLM-expanded candidates remain evidence grade D until independently validated or explicitly promoted, and they cannot be labeled observed language. They also cannot receive observed frequency, and non-default locales require native review before entering the core partition.

Openers like 'As a forward-thinking CFO...' do not reflect how real people prompt and distort study results. Prompts should include only the context a real person needs to get a useful answer, in concise, contextual, or naturally imperfect styles.