Creative Ads Explorer

Creative — Ads Explorer

Pump out a broad, diverse batch of ad hypotheses for A/B/n testing. This is the exploration stage: quantity + diversity beat polish. The goal is to cover the message space so the market tells you what works — not to perfect one ad.

When to use

• Launching paid social/display and need a test batch.
• An existing ad is fatiguing and you need fresh angles.
• Validating which message resonates before investing in production.

When NOT to use

• You already know the winning angle and just need it finished → creative-polish.
• You're merchandising a specific discount/bundle → creative-offer (sharper for that).
• No visual direction exists and brand look matters → lock it via creative-explore first.

Inputs (ask only what's missing)

• Product + core offer (what they get).
• Audience + their #1 pain or desire (the emotional hook).
• Platform: Meta, TikTok, Google Display, X, YouTube — sets aspect + length limits.
• Locked visual direction (from creative-explore) if available, so the batch stays on-brand.

The angle × format matrix

Build hypotheses by crossing angles with formats:

• Angles: pain, benefit, social-proof, urgency/scarcity, curiosity/pattern-interrupt, comparison/vs-incumbent, founder/origin story, transformation/before-after.
• Formats: static hero, hook-frame (thumb-stopper top third), UGC mock (handheld, casual), bold-text overlay, before/after split, problem→solution 2-panel.

Not every cell is worth filling — pick the combinations most likely to resonate for THIS audience.

Workflow

1. Pick the count (default 12, max 25). Choose the highest-potential angle×format cells.
2. For each: headline (≤7 words), subtext, visual prompt, and why it might win (the hypothesis being tested).
3. Kill near-duplicates aggressively — two ads testing the same idea waste spend.
4. Generate visuals with your image-generation tool at platform ratio (1024x1024 square, 1792x1024 landscape, or vertical for stories/TikTok). Save to ./creative/ads/-.png.
5. Output a ranked table; mark the top 3-5 to launch first and state what each one is testing.

Worked example

Product: AI meeting-notes app. Audience pain: "I forget action items." Platform: Meta feed (1:1).

| #  | angle      | format       | headline                   | tests hypothesis              | pri  |
|----|------------|--------------|----------------------------|-------------------------------|------|
| 1  | pain       | hook-frame   | "Forgot the action item again?" | pain-first beats feature  | HIGH |
| 2  | benefit    | static hero  | "Every decision, captured"  | calm-benefit framing          | HIGH |
| 3  | social     | UGC mock     | "My team stopped taking notes" | testimonial credibility    | HIGH |
| 4  | comparison | 2-panel      | "Otter vs us"               | switcher intent               | MED  |
| 5  | urgency    | bold-text    | "Your next meeting starts in 10" | urgency relevance        | LOW  |

Launch 1-3 first (distinct emotional levers); 4 if switcher traffic exists.

Quality bar

• Each ad tests a different hypothesis — you can name what you'd learn from each.
• Headlines respect platform character limits (Meta primary ≤125, headline ≤40, etc.).
• Text baked into images is short and exact; passed verbatim in the prompt.
• Top picks are justified, not arbitrary.

Common pitfalls

• Variation theater — 25 near-identical ads. Diversity of angle is the point.
• Fabricated proof — never invent stats, reviews, or testimonials. Flag any the user must supply real.
• Ignoring platform specs — wrong aspect/length gets rejected or cropped badly.
• Polishing too early — these are disposable hypotheses; only winners go to creative-polish.

Handoff

Winning ads (post-test or top-ranked) → creative-polish for production finish. On-brand consistency check → reuse the creative-explore stub.

Tooling

Needs a text-to-image tool to render. Without one, deliver the full matrix (headlines + prompts + hypotheses) and the user generates manually.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: DKeken
• Source: DKeken/codex-skills-alternative
• License: MIT

Install and usage instructions live in the source repository linked above.