--- name: nano-banana-ads description: Activates whenever the user asks to generate static ad images, product images, marketing visuals, or any image asset for Meta Ads, Instagram, or paid social. Uses Google's Nano Banana 2 image model via OpenRouter API. Always references brand-kit.md for colors, fonts, and style. Includes cost estimation, batch generation, and integration with the static-ad-designer skill for layout rules. --- # Nano Banana Ads Skill This skill handles image generation for static ad workflows using Google's Nano Banana 2 model (via OpenRouter). It's the production layer of Pillar 1's Static Ads Engine. The static-ad-designer skill handles strategy and layout rules. This skill handles the API calls. ## Activation Rules Activate when: - The user asks to generate ads, ad creatives, product images, or marketing visuals - The user invokes the static-ad-designer skill (which calls this skill internally) - The user runs any of the slash commands defined below - The user asks for image generation as part of any Vault workflow ## Prerequisites Before this skill can run, the user needs: 1. An OpenRouter API key stored as `OPENROUTER_API_KEY` environment variable, or already saved in Claude Code's API key store 2. A `brand-kit.md` file in the project root (use the brand-kit skill to create one) 3. Sufficient OpenRouter credits ($5+ recommended for a first batch) If any prerequisite is missing, prompt the user to set it up before continuing. --- ## API Configuration **Endpoint:** `https://openrouter.ai/api/v1/chat/completions` **Model ID:** `google/gemini-3.1-flash-image-preview` **Required headers:** ``` Authorization: Bearer ${OPENROUTER_API_KEY} Content-Type: application/json HTTP-Referer: https://highlyevolvedplant.com (or user's site) X-Title: HEP Vault Workflow ``` **Request body structure:** ```json { "model": "google/gemini-3.1-flash-image-preview", "messages": [ { "role": "user", "content": "[the prompt]" } ], "modalities": ["image", "text"] } ``` **Resolution defaults:** - 1K (1024x1024 or 1024x1280): For drafts and creative testing. ~$0.06-0.07 per image. - 2K (2048x2048 or 2048x2560): For finals and production. ~$0.08-0.10 per image. Default to 1K unless the user explicitly requests final/production quality. --- ## Generation Rules ### Always include Before sending any prompt, prepend brand context from `brand-kit.md`: ``` BRAND CONTEXT (from brand-kit.md): - Brand name: {brand_name} - Primary color: {primary_hex} - Secondary color: {secondary_hex} - Accent color: {accent_hex} - Font for headlines: {headline_font} - Photography style: {photography_style_description} - Voice/tone for any text in the image: {voice_description} - Banned words (must not appear): {banned_words_list} PROMPT: {user_prompt} ``` ### One image per call Make one API call per image, not batch requests. Nano Banana 2 returns one image per call. Batch requests don't work and will silently produce inconsistent results. ### Filename conventions Save images using the naming convention from brand-kit.md: ``` [brand]_[objective]_[hook]_[persona]_[style]_[version]_[date].png ``` Default save location: `./output/[concept-name].png` For aspect ratio variations, append the ratio: ``` [concept-name]_1x1.png (1:1 square) [concept-name]_4x5.png (4:5 vertical, Meta feed) [concept-name]_9x16.png (9:16 stories/reels) ``` ### Cost estimation before batch runs Before generating more than 5 images, show the user a cost estimate: ``` About to generate [N] images at [resolution] resolution. Estimated cost: $[N * price_per_image] Continue? (y/n) ``` After every 8 images in a batch, show running cost total. --- ## Slash Commands ### `/generate-ads-batch [matrix-file]` Read a creative matrix JSON file and generate all images in sequence. **Usage:** ``` /generate-ads-batch ./creative-matrix.json ``` **Workflow:** 1. Read the creative matrix file 2. Show cost estimate 3. Wait for confirmation 4. For each concept in the matrix: - Build the full prompt (brand context + concept-specific prompt + static-ad-designer rules) - Call Nano Banana 2 API - Save the returned image to `./output/[concept-name].png` - Generate both 1:1 and 4:5 versions 5. After every 8 images, show progress + running cost 6. At end, show contact sheet (grid of all generated images) ### `/regenerate [concept-name]` Regenerate a single image with new instructions. **Usage:** ``` /regenerate vitapure_sales_problem_sarah_minimal_v1 > "Make the background warmer, less stark white. Add subtle paper texture." ``` ### `/upscale-finals` Take user-selected concepts and regenerate them at 2K resolution for production use. **Usage:** ``` /upscale-finals vitapure_sales_problem_sarah_minimal_v1, vitapure_sales_benefit_marcus_lifestyle_v1 ``` Saves to `./output/final/` directory. ### `/generate-variations [concept-name]` Generate 3 variations of a single concept with the same hook and persona but different visual treatment. **Usage:** ``` /generate-variations vitapure_sales_problem_sarah_minimal_v1 ``` --- ## Common Issues and Fixes ### "Text in the image is illegible" Add to the prompt: ``` TEXT REQUIREMENTS: - All text must be at least 48px equivalent - Add a solid or semi-transparent panel behind any headline text - Use brand headline font: {headline_font} - Maintain WCAG AA contrast (4.5:1 minimum) between text and its background ``` ### "Colors don't match brand" Strengthen the color enforcement: ``` COLOR REQUIREMENTS (strict): - Background: must be exactly {background_hex} - Headline color: must be exactly {primary_hex} - CTA button: must be exactly {accent_hex} - Do not use any colors not in the brand palette ``` ### "Image looks like generic AI slop" Update the static-ad-designer skill (not this one) — slop is usually a strategy/composition problem, not a generation problem. But for one-off fixes, add: ``` ANTI-SLOP REQUIREMENTS: - No stock-photo aesthetic (no overly polished, generic smiling people) - No cluttered layouts (max 3 visual elements per frame) - No tiny text that's unreadable in a feed scroll - Asymmetric composition preferred over centered - Real-world textures over flat gradients ``` ### "Character or product looks different across batch" For multi-image batches that need consistency: ``` CONSISTENCY REQUIREMENTS: - Character description (use exact text in every prompt): {character_description} - Product reference: attach the same reference image to every call - Lighting direction: {direction} (must be identical across all images) - Color palette: pulled from brand-kit.md (no deviations) ``` --- ## Cross-Skill Integration When the static-ad-designer skill calls this skill: 1. Static ad designer provides the strategic prompt (hook, persona, visual concept) 2. This skill prepends brand context from brand-kit.md 3. This skill makes the API call 4. This skill saves the image with the correct naming convention 5. This skill returns the file path back to static-ad-designer When the animation-ads skill calls this skill: - Same flow, but generates frames as part of a 3x3 grid for storyboard consistency - See animation-ads-skill.md for grid prompt structure When the creative-strategist skill (Pillar 3) recommends new concepts: - It outputs a new creative matrix - User runs `/generate-ads-batch [new-matrix.json]` and the loop continues --- ## Output Format After generation, return: ``` Generated [N] images. Saved to: ./output/ Cost: $[total] Time: [seconds] Concepts generated: - [concept-name-1] (1:1, 4:5) - [concept-name-2] (1:1, 4:5) ... Next steps: 1. Review images in ./output/ 2. Pick top 10-15 winners 3. Run /generate-variations on winners for additional testing 4. Run /upscale-finals on final selections for production use 5. Upload to Meta Ads Creative Library 6. Use Pillar 2 (Bulk Uploader) to push to Ads Manager ``` --- ## Notes - This skill assumes Nano Banana 2 (model ID `google/gemini-3.1-flash-image-preview`). Update the model ID when Nano Banana 3 launches. The API structure is unlikely to change. - OpenRouter pricing fluctuates monthly. Verify current pricing at openrouter.ai/google/gemini-3.1-flash-image-preview before quoting costs to clients. - For agency use with multiple client accounts, use a separate OpenRouter project per client to track costs separately.