Ad Creative Design

Designs ad creatives that are platform-native (correct dimensions, safe zones, text overlay policy, character limits), industry-precise (visual DNA per industry × ad objective), and prompt-sanitized (no platform names / UI artifacts / fake ad elements leak into the image generator). The skill lays out an 8-platform × 21-format matrix, a six-step workflow, a self-check gate, and a banned-words system with a positive-preservation companion rule.

Engagement Principles

These rules apply across every Section. Read before acting.

1. No generic one-size-fits-all creatives — every creative is locked to a specific (platform, format, objective, industry) tuple. No tuple = no work.
2. Hard Constraints first — Section 0 (Platform Dimensions / Safe Zones / Text Overlay / Ad Copy Char Limits) is loaded at Step 0 and re-validated at the Self-Check Gate before delivery. Violations like "text on Google Display" or "CTA in TikTok bottom 18%" cause platform takedowns or render the ad invisible.
3. Brand Name & User Copy Fidelity is MUST-level — every brand name, slogan, CTA text, and price/offer the user specified must appear verbatim in the deliverable. Never paraphrase, translate, or substitute user-supplied copy.
4. Industry × Objective drives composition + color — pick the composition pattern and color strategy from the cross-decision matrix (Section 8 / Section 9), not from generic style intuition.
5. Platform Style Profiles are not stylistic suggestions — they are constraints. TikTok requires anti-ad aesthetic AND complete brand/product/CTA info; LinkedIn requires ≥60pt fonts; Pinterest requires aspirational scenes; Google Display requires zero text overlay.
6. Prompt Banned Words are filtered out of the image generator prompt — platform names, ad-UI terms, brand-tool names never reach the model. They are used internally for routing and rule selection only.
7. Required Positive Instruction protects native packaging text — "no text overlay" must not delete printed text on the product itself. Always pair the banned-words filter with the positive preservation instruction (references/banned-words.md §Required Positive Instruction).
8. Match the user's language — respond in English if the user writes English, in Chinese if Chinese. Never switch language unprompted.
9. Use <suggestion> for option sets — when proposing platform / objective / industry / asset choices, wrap them in <suggestion> tags. Never ask multiple questions at once.

Execution Procedure