Claude Skill

ad-creative-generation

Generate on-brand ad creatives — visuals + copy — for Google, Meta (Facebook / Instagram), and other paid platforms via the Hyper MCP. Extracts brand identity from a website, writes ad copy variants, and produces brand-consistent images using reference-based image generation. Use

LLM Mart · 0 points · 3 views 45 listing impressions 0 install-command copies
Download hyperfx-ai-marketing-skills-skills_ad-creative-generation-70b0b30.zip · 13 KB

Install

skills CLI npx skills add https://github.com/hyperfx-ai/marketing-skills/tree/main/skills/ad-creative-generation
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install hyperfx-ai-marketing-skills@llmmart
Git git clone https://github.com/hyperfx-ai/marketing-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole hyperfx-ai/marketing-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Ad Creative Generation

Generate ad creatives — both visuals and copy — that match a brand's identity. This skill orchestrates brand extraction, copywriting, and image generation into a single workflow.

Requirements

This skill assumes the Hyper MCP is connected to your agent so the tools below are available. Brand extraction also requires Firecrawl to be configured under your Hyper integrations.

How to run the tools in this skill

Every tool in this skill is named by its canonical tool name. Run it with the call your surface gives you:

Surface Find a tool Run it
Hyper chat Call the tool by its name; search_tools(queries=["<provider> <action>"]) shows its parameters <name>(...)
MCP client (Claude, Cursor, Codex, ChatGPT) search("<what you want to do>"), then describe("<name>") call_read("<name>", {...}), call_write or call_destructive: the one that describe names in call_with
Hyper CLI hyperai search "<what you want to do>", then hyperai describe <name> hyperai call <name> --json '{...}'

If a tool is not found: in Hyper chat, find its toolkit id with discover_toolkits and pass it to enable_toolkit. On an MCP client or the CLI, tell the user which integration to connect in Hyper.

Tool surface

Group Tools
Brand extraction firecrawl_branding_extract
Competitor ads marketing_queries_start, marketing_queries_continue (stream ads.meta_library)
Image generation (default) images_generate (model="gpt-image-2")
Image generation (text-heavy) images_generate (model="nano-banana-pro")
Image generation (photoreal product shots) images_generate (model="seedream-4.5")

For deeper image-tool selection guidance see the image-generation skill. To turn finished creatives into running campaigns see google-ads and meta-ads.

Out of scope

  • Picking the right image model when the task isn't ad creative — use image-generation.
  • Creating, launching, or budgeting actual ad campaigns — use google-ads, meta-ads, tiktok-ads, pinterest-ads, or amazon-ads.
  • Searching live competitor ads — use meta-ads-library. To read the Meta ads of a known competitor by its domain or Meta page in one query, use marketing_queries_start with the ads.meta_library stream; a repeated query within a day reads the stored run at no cost.

Critical rules

  • Always extract branding first before generating visuals for a website/brand.
  • Always pass both logo and screenshot as reference images — SVG logos are auto-converted to PNG.
  • DO NOT display image URLs — they are automatically shown in chat.
  • Read brand.screenshot.description to understand the product — never guess from the company name.
  • Default to logo + headline + product screenshot for SaaS / product ads unless the user requests a different style.
  • Use model="gpt-image-2" for the initial ad creative by default — with reference images for branded work, or without for loose first-pass concepts.
  • Do not default to model="nano-banana-pro" for the first creative pass — use it when the user explicitly asks or when readable text inside the image is the main requirement.
  • Set aspect_ratio to the placement — 1:1 (feed), 9:16 (story), 16:9 (banner) — and quality ("draft" | "standard" | "high"); images_generate takes those, not raw pixel sizes.
  • Respect character limits — Google RSA headlines are 30 chars, descriptions 90 chars; Meta primary text is 125 chars visible.
  • Change one variable per variant so test results are attributable.
  • Match aspect ratio to placement — feed is 1:1, story is 9:16 (use 1024x1536), banner is 16:9 (use 1536x1024).

Routing Table

Based on what the user needs, read the appropriate reference file:

User Need Reference
"Extract branding from this site" / brand colors, logo references/brand-extraction.md
"Write ad copy" / headlines / variants / hooks references/ad-copy-frameworks.md
Google RSAs / display ads / Performance Max assets references/google-ads-creatives.md
Meta / Facebook / Instagram ads / carousel / stories references/meta-ads-creatives.md

When the task spans multiple areas (e.g., "create ad creatives for this website"), follow the full workflow below.

Core Workflow

Phase 1: Brand Extraction

Call firecrawl_branding_extract with the website URL. This single call returns branding data, a saved logo, and a website screenshot:

brand = firecrawl_branding_extract(url="https://example.com")

The result contains:

  • brand.logo.file_id — logo image to use as reference
  • brand.screenshot.file_id — website screenshot to use as reference
  • brand.branding — colors, typography, spacing, personality

Both logo and screenshot are saved images ready to pass as reference_images to image generation tools. SVG logos are automatically converted to PNG, so always include the logo — no need to filter by format.

Read references/brand-extraction.md only when the site gives no logo, colors or screenshot.

Phase 2: Ad Copy Generation

Using the brand personality, tone, and value proposition from the extraction, write ad copy variants. Structure copy by what's being tested:

  1. Hook variants — different opening angles (number-led, question, pain-point, benefit-first)
  2. Body variants — different messaging (social proof, feature highlight, urgency)
  3. CTA variants — different calls to action (Learn More, Shop Now, Get Started, Try Free)

Change only one element per variant so results can be attributed to specific changes.

The platform file of Phase 4 holds the copy rules of its placements. Read references/ad-copy-frameworks.md only for copy with no placement.

Phase 3: Visual Creative Generation

Generate ad images using brand assets as references. Always pass both logo and screenshot as reference images — SVG logos are automatically converted to PNG by the branding tool.

Read brand.screenshot.description to understand what the product actually does and what the UI looks like. Do not guess from the company name.

Preferred first pass: images_generate with model="gpt-image-2"

Default approach (best practice for SaaS / product ads)

By default, a strong ad creative has three layers: brand (logo), copy (headline), and product (realistic screenshot showing the core value prop). This is the recommended starting point when the user hasn't specified a creative direction. Use model="gpt-image-2" for this initial composition unless the user explicitly asks for another model or the image is primarily a text-rendering task.

images_generate(
    requests=[{
        "prompt": (
            "Social media ad creative for [company name]. "
            "Top: the [company] logo. "
            "Headline: '[headline from ad copy phase or site hero text]'. "
            "Below the headline: a clean, realistic product screenshot of [describe the actual UI "
            "based on brand.screenshot.description]. "
            "Match the brand style from the references."
        ),
        "reference_images": [brand.logo.file_id, brand.screenshot.file_id]
    }],
    model="gpt-image-2",
    aspect_ratio="1:1",
    quality="high"
)

Following user direction

If the user asks for a specific creative style (lifestyle imagery, abstract, illustration, people using the product, etc.), follow their direction. The defaults above are a starting point, not a constraint. The reference images still supply brand consistency regardless of creative direction.

Things to avoid (unless the user specifically asks)

  • Made-up visual elements (robot mascots, random platform logos, abstract graphics)
  • UI chrome (buttons, nav bars, form inputs) — ad platforms add their own CTAs
  • Guessing what the product does from the company name
  • Generic marketing clip art

Visual Tool Selection

Scenario Tool Why
On-brand creative (default) images_generate (model="gpt-image-2") with logo + screenshot refs Best default for the first branded concept
Text-heavy creative (headlines in image) images_generate (model="nano-banana-pro") Use only when text rendering inside the image is the main requirement
Quick ideation / concept exploration images_generate (model="gpt-image-2") Fast first-pass concepting with no references needed
Iterative refinement of existing image images_generate (model="nano-banana") with the image as a reference Edit a specific generated image
Photoreal product shots / material detail images_generate (model="seedream-4.5") Strong fabric/texture/spatial depth

Platform Quick Reference

Platform Format Image Size Key Limits
Meta Feed 1:1 1080x1080 Primary text: 125 chars visible
Meta Story 9:16 1080x1920 Full screen, 15s max
Meta Carousel 1:1 1080x1080 Up to 10 cards
Google RSA N/A (text only) N/A 15 headlines (30 chars), 4 descriptions (90 chars)
Google Display Various 1200x628 Responsive display ads

For full platform specs, read references/google-ads-creatives.md or references/meta-ads-creatives.md.

Ad Copy Variant Quick Reference

When generating variants from existing top performers:

  1. Identify the winning elements: hook type, CTA, messaging angle, format
  2. Generate variants that preserve winning elements while changing one variable
  3. Group variants by what's being tested (hook, body, CTA)
  4. Pair variants for A/B testing — each pair should isolate one variable

For detailed frameworks, read references/ad-copy-frameworks.md.

  • Follow user direction — if the user wants lifestyle, illustration, abstract, or any other style, follow their lead. The reference images still supply brand consistency regardless of creative direction.
Files (marketing-skills)
  • references
    • ad-copy-frameworks.md 5.7 KB
      Loading…
    • brand-extraction.md 3.7 KB
      Loading…
    • google-ads-creatives.md 5.5 KB
      Loading…
    • meta-ads-creatives.md 6.8 KB
      Loading…
  • SKILL.md 10.6 KB
    ---
    name: ad-creative-generation
    description: Generate on-brand ad creatives — visuals + copy — for Google, Meta (Facebook / Instagram), and other paid platforms via the Hyper MCP. Extracts brand identity from a website, writes ad copy variants, and produces brand-consistent images using reference-based image generation. Use when the user asks for ad creative, ad copy variants, RSA headlines, Meta ad creative, display ads, carousel ads, or A/B test variants.
    metadata:
      version: 1.1.0
    requires_toolkits:
      - image_gen
    icon: image_gen
    short_description: Brand extraction, ad copy variants, and on-brand images for Google and Meta placements.
    ---
    
    # Ad Creative Generation
    
    Generate ad creatives — both visuals and copy — that match a brand's identity. This skill orchestrates brand extraction, copywriting, and image generation into a single workflow.
    
    ## Requirements
    
    This skill assumes the [Hyper MCP](https://app.hyperfx.ai/mcp) is connected to your agent so the tools below are available. Brand extraction also requires Firecrawl to be configured under your Hyper integrations.
    
    ### How to run the tools in this skill
    
    Every tool in this skill is named by its canonical tool name. Run it with the call your surface gives you:
    
    | Surface | Find a tool | Run it |
    | --- | --- | --- |
    | Hyper chat | Call the tool by its name; `search_tools(queries=["<provider> <action>"])` shows its parameters | `<name>(...)` |
    | MCP client (Claude, Cursor, Codex, ChatGPT) | `search("<what you want to do>")`, then `describe("<name>")` | `call_read("<name>", {...})`, `call_write` or `call_destructive`: the one that `describe` names in `call_with` |
    | Hyper CLI | `hyperai search "<what you want to do>"`, then `hyperai describe <name>` | `hyperai call <name> --json '{...}'` |
    
    If a tool is not found: in Hyper chat, find its toolkit id with `discover_toolkits` and pass it to `enable_toolkit`. On an MCP client or the CLI, tell the user which integration to connect in Hyper.
    
    ## Tool surface
    
    | Group | Tools |
    |-------|-------|
    | Brand extraction | `firecrawl_branding_extract` |
    | Competitor ads | `marketing_queries_start`, `marketing_queries_continue` (stream `ads.meta_library`) |
    | Image generation (default) | `images_generate` (`model="gpt-image-2"`) |
    | Image generation (text-heavy) | `images_generate` (`model="nano-banana-pro"`) |
    | Image generation (photoreal product shots) | `images_generate` (`model="seedream-4.5"`) |
    
    For deeper image-tool selection guidance see the `image-generation` skill. To turn finished creatives into running campaigns see `google-ads` and `meta-ads`.
    
    ## Out of scope
    
    - Picking the right image model when the task isn't ad creative — use `image-generation`.
    - Creating, launching, or budgeting actual ad campaigns — use `google-ads`, `meta-ads`, `tiktok-ads`, `pinterest-ads`, or `amazon-ads`.
    - Searching live competitor ads — use `meta-ads-library`. To read the Meta ads of a known competitor by its domain or Meta page in one query, use `marketing_queries_start` with the `ads.meta_library` stream; a repeated query within a day reads the stored run at no cost.
    
    ## Critical rules
    
    - **Always extract branding first** before generating visuals for a website/brand.
    - **Always pass both logo and screenshot** as reference images — SVG logos are auto-converted to PNG.
    - **DO NOT display image URLs** — they are automatically shown in chat.
    - **Read `brand.screenshot.description`** to understand the product — never guess from the company name.
    - **Default to logo + headline + product screenshot** for SaaS / product ads unless the user requests a different style.
    - **Use `model="gpt-image-2"` for the initial ad creative by default** — with reference images for branded work, or without for loose first-pass concepts.
    - **Do not default to `model="nano-banana-pro"` for the first creative pass** — use it when the user explicitly asks or when readable text inside the image is the main requirement.
    - **Set `aspect_ratio` to the placement** — `1:1` (feed), `9:16` (story), `16:9` (banner) — and `quality` (`"draft" | "standard" | "high"`); `images_generate` takes those, not raw pixel sizes.
    - **Respect character limits** — Google RSA headlines are 30 chars, descriptions 90 chars; Meta primary text is 125 chars visible.
    - **Change one variable per variant** so test results are attributable.
    - **Match aspect ratio to placement** — feed is 1:1, story is 9:16 (use `1024x1536`), banner is 16:9 (use `1536x1024`).
    
    ## Routing Table
    
    Based on what the user needs, read the appropriate reference file:
    
    | User Need | Reference |
    |-----------|-----------|
    | "Extract branding from this site" / brand colors, logo | `references/brand-extraction.md` |
    | "Write ad copy" / headlines / variants / hooks | `references/ad-copy-frameworks.md` |
    | Google RSAs / display ads / Performance Max assets | `references/google-ads-creatives.md` |
    | Meta / Facebook / Instagram ads / carousel / stories | `references/meta-ads-creatives.md` |
    
    When the task spans multiple areas (e.g., "create ad creatives for this website"), follow the full workflow below.
    
    ## Core Workflow
    
    ### Phase 1: Brand Extraction
    
    Call `firecrawl_branding_extract` with the website URL. This single call returns branding data, a saved logo, and a website screenshot:
    
    ```python
    brand = firecrawl_branding_extract(url="https://example.com")
    ```
    
    The result contains:
    - `brand.logo.file_id` — logo image to use as reference
    - `brand.screenshot.file_id` — website screenshot to use as reference
    - `brand.branding` — colors, typography, spacing, personality
    
    Both `logo` and `screenshot` are saved images ready to pass as `reference_images` to image generation tools. SVG logos are automatically converted to PNG, so always include the logo — no need to filter by format.
    
    **Read `references/brand-extraction.md` only when the site gives no logo, colors or screenshot.**
    
    ### Phase 2: Ad Copy Generation
    
    Using the brand personality, tone, and value proposition from the extraction, write ad copy variants. Structure copy by what's being tested:
    
    1. **Hook variants** — different opening angles (number-led, question, pain-point, benefit-first)
    2. **Body variants** — different messaging (social proof, feature highlight, urgency)
    3. **CTA variants** — different calls to action (Learn More, Shop Now, Get Started, Try Free)
    
    Change only one element per variant so results can be attributed to specific changes.
    
    **The platform file of Phase 4 holds the copy rules of its placements. Read `references/ad-copy-frameworks.md` only for copy with no placement.**
    
    ### Phase 3: Visual Creative Generation
    
    Generate ad images using brand assets as references. Always pass **both logo and screenshot** as reference images — SVG logos are automatically converted to PNG by the branding tool.
    
    **Read `brand.screenshot.description`** to understand what the product actually does and what the UI looks like. Do not guess from the company name.
    
    **Preferred first pass: `images_generate` with `model="gpt-image-2"`**
    
    #### Default approach (best practice for SaaS / product ads)
    
    By default, a strong ad creative has three layers: **brand** (logo), **copy** (headline), and **product** (realistic screenshot showing the core value prop). This is the recommended starting point when the user hasn't specified a creative direction. Use `model="gpt-image-2"` for this initial composition unless the user explicitly asks for another model or the image is primarily a text-rendering task.
    
    ```python
    images_generate(
        requests=[{
            "prompt": (
                "Social media ad creative for [company name]. "
                "Top: the [company] logo. "
                "Headline: '[headline from ad copy phase or site hero text]'. "
                "Below the headline: a clean, realistic product screenshot of [describe the actual UI "
                "based on brand.screenshot.description]. "
                "Match the brand style from the references."
            ),
            "reference_images": [brand.logo.file_id, brand.screenshot.file_id]
        }],
        model="gpt-image-2",
        aspect_ratio="1:1",
        quality="high"
    )
    ```
    
    #### Following user direction
    
    If the user asks for a specific creative style (lifestyle imagery, abstract, illustration, people using the product, etc.), follow their direction. The defaults above are a starting point, not a constraint. The reference images still supply brand consistency regardless of creative direction.
    
    #### Things to avoid (unless the user specifically asks)
    
    - Made-up visual elements (robot mascots, random platform logos, abstract graphics)
    - UI chrome (buttons, nav bars, form inputs) — ad platforms add their own CTAs
    - Guessing what the product does from the company name
    - Generic marketing clip art
    
    ## Visual Tool Selection
    
    | Scenario | Tool | Why |
    |----------|------|-----|
    | On-brand creative (default) | `images_generate` (`model="gpt-image-2"`) with logo + screenshot refs | Best default for the first branded concept |
    | Text-heavy creative (headlines in image) | `images_generate` (`model="nano-banana-pro"`) | Use only when text rendering inside the image is the main requirement |
    | Quick ideation / concept exploration | `images_generate` (`model="gpt-image-2"`) | Fast first-pass concepting with no references needed |
    | Iterative refinement of existing image | `images_generate` (`model="nano-banana"`) with the image as a reference | Edit a specific generated image |
    | Photoreal product shots / material detail | `images_generate` (`model="seedream-4.5"`) | Strong fabric/texture/spatial depth |
    
    ## Platform Quick Reference
    
    | Platform | Format | Image Size | Key Limits |
    |----------|--------|-----------|------------|
    | Meta Feed | 1:1 | 1080x1080 | Primary text: 125 chars visible |
    | Meta Story | 9:16 | 1080x1920 | Full screen, 15s max |
    | Meta Carousel | 1:1 | 1080x1080 | Up to 10 cards |
    | Google RSA | N/A (text only) | N/A | 15 headlines (30 chars), 4 descriptions (90 chars) |
    | Google Display | Various | 1200x628 | Responsive display ads |
    
    **For full platform specs, read `references/google-ads-creatives.md` or `references/meta-ads-creatives.md`.**
    
    ## Ad Copy Variant Quick Reference
    
    When generating variants from existing top performers:
    
    1. Identify the winning elements: hook type, CTA, messaging angle, format
    2. Generate variants that preserve winning elements while changing one variable
    3. Group variants by what's being tested (hook, body, CTA)
    4. Pair variants for A/B testing — each pair should isolate one variable
    
    **For detailed frameworks, read `references/ad-copy-frameworks.md`.**
    
    - **Follow user direction** — if the user wants lifestyle, illustration, abstract, or any other style, follow their lead. The reference images still supply brand consistency regardless of creative direction.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.