Realistic Prompt Generation

Write authentic prompts without manufacturing recommendation opportunities.

This skill inherits the ethical floor from skills/ETHICS.md. It enforces anti-hallucination and evidence-bound language. Anti-spray and human-send are not applicable.

Hard context boundary

For unaided generation, accept only:

• anonymized segment and role labels;
• approved jobs, constraints, information acts, journey states, and locales;
• short source-language fragments safe under the permitted-data rules;
• evidence IDs and grades;
• prompt_architecture.json;
• style and turn-form requirements.

Do not accept or inspect:

• target brands, products, domains, people, slogans, campaign terms, proprietary categories, or flattering claims;
• current AI answers, rankings, mentions, citations, gaps, or target pages;
• the contamination register itself.

If those fields appear, stop unaided generation and request a sanitized blind_design_brief.json. When subagents or fresh sessions are available, generate in a fresh context that receives only the blind brief and architecture.

The exception is an explicitly separate B0 aided pass. It may receive only the target aliases needed by approved B0 cells.

Preserve the canonical intent cell

For each architecture cell, hold constant:

• underlying job;
• journey state and information act;
• material constraints;
• persona/role and locale;
• proximity band;
• expected answer kind.

Create two core variants by default:

1. the closest natural rendering of observed language;
2. a natural paraphrase that preserves the same intent.

Use additional variants only for a wording-sensitivity pilot or rotating discovery. Do not create a full style × persona × locale grid.