Static Ad Generator Skill for AI Agents

Just ask

Say the word, it runs.

The skill fires on natural requests like these. No commands to memorize.

“Generate 15 static ad concepts for my brand”“Turn my reviews into ad creative”“Make me a us vs them ad”“I need more ad variations to test”

The fifteen templates

One brand. Fifteen structures.

Every ad below is the same brand, the same research and the same batch, run through a different layout. That is the whole idea: template diversity is angle diversity, and the winner is rarely the one you would have picked by hand. Each tile names the source its copy came from.

Headline statement

One claim carries the whole ad. Product underneath, nothing else in the frame.

Sourced from: Brand brief, Angles

Us vs. them

Two columns, four rows. The old way greyed out, the brand in full colour.

Sourced from: Store feature list, no competitor named

Stat callout

One defensible number at sixty percent of the visual weight, its denominator kept.

Sourced from: Live App Store rating data

Review card

A real review, verbatim, styled as the native review UI its buyers actually read.

Sourced from: Brand brief, Voice of Customer

Testimonial stack

Several quotes stacked, each covering a different objection. Initial avatars, never a generated face.

Sourced from: The brand's own store screenshots

Before / after

Two panels and an arrow. Workflow and ritual transformations, not regulated body claims.

Sourced from: Brand brief, Before to After

Problem / solution

Tension on top in the customer's exact words, the product as the answer below.

Sourced from: Brand brief, Objections

Founder message

A plain note, no product glamour shot. Drafted for sign-off, never invented in first person.

Sourced from: Dossier, community-sourced trail database

Feature spotlight

Product centred, callouts radiating out. The components buyers ask about, not the full feature list.

Sourced from: App Store feature list

Press mention

Real coverage only, outlet names set as plain wordmarks rather than borrowed logos.

Sourced from: The Cut, New York Magazine

Lifestyle hero

One photograph does the work. Five to eight words of identity copy, and the logo.

Sourced from: The brand's own slogan

Numbered list

Four rows, four distinct angles: pain, outcome, proof, differentiator.

Sourced from: Trail detail fields and filters

FAQ card

A real objection, phrased the way customers phrase it, answered in one line.

Sourced from: Brand brief, Objections and Rebuttals

Competitor callout

One factual axis of difference. The name in plain text, never their logo or packaging.

Sourced from: Brand brief, Why Us Not Them

Origin story

Longer copy than any other format. The specific moment that started it is the credibility.

Sourced from: Dossier, 30M+ contributors

Rendered for AllTrails in a single run: fifteen concepts, gpt-image-2 at high quality, 4:5, on brand.

The pipeline

What the skill does for you.

1
Build the source bank
The skill reads your brand brief, dossier, product catalog and app-store data, then sorts the raw material into buckets: verbatim customer quotes, proof numbers with their denominators, pain language, objections, named competitors, press mentions, the founder's real story. It asks once whether you have winning ads or reviews of your own to add.
2
Spread the batch across all fifteen templates
A batch clustered on three favourite layouts is three ads in fourteen costumes. Fifteen concepts means one per template; thirty means two each. If a template has no grounded material for your brand, it gets skipped and listed, never filled with a plausible-sounding number.
3
Write the concept sheet
Every concept lands as a card with its template, headline, body, visual and the source it came from, in text, before a credit is spent. This is the real approval gate: copy edits are free, a regenerated image is not.
4
Render with the copy burned in
One gpt-image-2 call per approved concept at high quality, on-brand, with your logo and brand-kit board attached and the real product passed as a reference. Advibly has no text-overlay step by design, so every quoted line is written into the prompt exactly as it should read.
5
Read every render against its card
Is each quoted line present and spelled right, is the product the real product, did the model add a fake badge, an award or a second logo. Misses get flagged honestly and rerolled with fewer words or a simpler layout.
6
Deliver the index
A table of concept number, template, headline, grounding and image URL, so you can drop concepts straight into your ad platform and trace every claim back to its source. Plus what was skipped and why.

Six steps. You approve one sheet.

Every concept arrives as text first. Edit the copy for free, approve it, and the renders run.

Start creating

Grounded copy

Every line has a receipt.

These four carry the claims generated ads usually make up: a rating, a review, a press quote, a competitor. Each one came from AllTrails’ own data and is cited on its concept card before anything is rendered.

Numbers keep their denominator: 4.9 from 1,032,380 App Store ratings, not just 4.9
Reviews are quoted word for word, never tidied into marketing voice
Press is real coverage only, with the outlet set as a plain wordmark
A competitor is named in plain text only when the brief names it, never its logo

Start creating Install the skill

Stat callout

Review card

Press mention

Competitor callout

The model stack

Image models that write their own copy.

Advibly has no text-overlay step, so every model here renders the headline into the image itself. GPT Image 2 at high quality is the default for every concept, and Seedream 5 Pro steps in when a dense layout keeps merging rows after two rerolls. Ask for GPT Image 2.5 Sunburst, GPT Image 2.5 Flare or Nano Banana 2 and the agent renders the batch on that instead.

What you get

Fifteen finished 4:5 static ads, one per template, ready for Meta
A concept sheet you approve before anything is rendered
A source citation on every headline, body line and stat
An index table with template, copy, grounding and URL per concept

Start creating Read the source

GPT Image 2.5 SunburstNew

The sharpest type, for the ads you plan to scale

GPT Image 2.5 FlareNew

The fastest pass over a big batch

Seedream 5 ProNew

Fallback for dense multi-row layouts

GPT Image 2

The default: every concept, copy burned in

Nano Banana 2

Photo-led layouts, up to 4K

Install

Add it in seconds.

Install the skill, then make sure the Advibly MCP is connected so it has tools to call. Works with every skills-capable agent.

Claude Code

Claude

Cursor

Codex

VS Code

OpenClaw

Hermes

ChatGPT

Connect the Advibly MCP first

One click, no terminal

Download .skill Download .zip

Drop the .skill file into Claude → Settings → Skills, or unzip the folder into ~/.claude/skills.

One command (any agent)

npx skills add advibly/skills -s static-ads

Works with Claude Code, Claude Desktop, Cursor, and any agent that supports skills. Pulls straight from GitHub.

Manual (Claude Code)

git clone https://github.com/advibly/skills.git
cp -R skills/static-ads ~/.claude/skills/

Or in Claude Desktop: Settings → Skills → Install and select the skill folder.

Bundled references

The playbook.

Each skill ships with reference docs the agent loads on demand, so the craft is baked in.

templates.md

All fifteen templates: structure, copy slot, examples per business type, where to source it, and the exact prompt skeleton.

grounding.md

The sourcing and compliance rules: verbatim quotes, real stats, real people's faces, competitor names, regulated claims.

Honest limits

What to watch for.

A template with no grounded material is skipped and named, not filled in. A brand with no brief, no reviews and no press gets fewer than fifteen concepts.
Dense layouts (comparison tables, numbered lists, testimonial stacks) can merge rows. The fix is four rows instead of five, then seedream-5-pro.
No generated likeness of a real person, ever. Founders and named reviewers get their real photo as a reference, initials avatars, or a faceless frame.
Competitor names are set in plain text only, and only when the brief names them. Never their logo or packaging.
Before-and-after is regulated in health, finance and beauty. It defaults to workflow and ritual transformations unless you confirm the platform policy.

Questions

Before you install it.

What is the Static Ads skill?

An agent skill that turns one brand into a batch of static image ads spread across fifteen proven layouts: headline statement, us vs. them, stat callout, review card, founder message, FAQ card and the rest. Each concept is rendered with gpt-image-2 with its copy written into the image, and every line cites where it came from.

What do I have to give the agent?

A brand that is already onboarded in Advibly and how many concepts you want. The default is fifteen, one per template, at 4:5 for Meta and Instagram feed; ask for 1:1, 9:16 or 16:9 instead. It asks once whether you have winning ads or reviews of your own, because real reviews with names beat anything scraped.

How long does a batch take?

The concept sheet is the part that needs you. It is written in text, costs nothing, and you approve it before a credit is spent. After that every render is an independent call, so they run in parallel, and you get an index of template, headline, source and image URL for each ad.

What does it cost?

The skill itself is free and open source. Running it spends your existing Advibly credits on the images it renders, at no extra skill cost, and every generation shows its credit cost before it runs.

Which agents can run it?

Any agent that supports installable skills and the Model Context Protocol: Claude Code, Claude, Cursor, Codex, VS Code, OpenClaw, Hermes and ChatGPT. Install with one npx command, then connect the Advibly MCP so the skill has tools to call.

What if my brand has no reviews or press yet?

Those templates are skipped and listed with the reason, never filled with a plausible-sounding number or an outlet that never covered you. A brand with no brief, no reviews and no press gets fewer than fifteen concepts, and every one it does get is honest.